The Impact of Sampling Schemes on the Site Frequency Spectrum in Nonequilibrium Subdivided Populations

Transcription

1 Copyright Ó 2009 by the Genetics Society of America DOI: /genetics The Impact of Sampling Schemes on the Site Frequency Spectrum in Nonequilibrium Subdivided Populations Thomas Städler,*,1 Bernhard Haubold, Carlos Merino, Wolfgang Stephan and Peter Pfaffelhuber *Institute of Integrative Biology, Plant Ecological Genetics, ETH Zurich, CH-8092 Zurich, Switzerland, Department of Evolutionary Genetics, Max-Planck-Institute for Evolutionary Biology, D Plön, Germany, Department of Biology II, Section of Evolutionary Biology, University of Munich, D Planegg-Martinsried, Germany and Faculty of Mathematics and Physics, University of Freiburg, D Freiburg, Germany Manuscript received August 6, 2008 Accepted for publication February 17, 2009 ABSTRACT Using coalescent simulations, we study the impact of three different sampling schemes on patterns of neutral diversity in structured populations. Specifically, we are interested in two summary statistics based on the site frequency spectrum as a function of migration rate, demographic history of the entire substructured population (including timing and magnitude of specieswide expansions), and the sampling scheme. Using simulations implementing both finite-island and two-dimensional stepping-stone spatial structure, we demonstrate strong effects of the sampling scheme on Tajima s D (D T ) and Fu and Li s D (D FL ) statistics, particularly under specieswide (range) expansions. Pooled samples yield average D T and D FL values that are generally intermediate between those of local and scattered samples. Local samples (and to a lesser extent, pooled samples) are influenced by local, rapid coalescence events in the underlying coalescent process. These processes result in lower proportions of external branch lengths and hence lower proportions of singletons, explaining our finding that the sampling scheme affects D FL more than it does D T. Under specieswide expansion scenarios, these effects of spatial sampling may persist up to very high levels of gene flow (Nm. 25), implying that local samples cannot be regarded as being drawn from a panmictic population. Importantly, many data sets on humans, Drosophila, and plants contain signatures of specieswide expansions and effects of sampling scheme that are predicted by our simulation results. This suggests that validating the assumption of panmixia is crucial if robust demographic inferences are to be made from local or pooled samples. However, future studies should consider adopting a framework that explicitly accounts for the genealogical effects of population subdivision and empirical sampling schemes. ALTHOUGH most (if not all) species are characterized by various degrees of population subdivision, population geneticists continue to employ models of panmictic populations of constant size (the neutral equilibrium, NE, model) when testing for selection and/ or demographic changes through time. Motivated by empirical data sets that were found to be incompatible with expectations under the NE model, recent largescale studies have employed coalescent simulations of demographic changes such as bottlenecks and/or population expansions in attempts to estimate parameters of such (arguably) biologically more realistic scenarios (e.g., Marth et al. 2004; Nordborg et al. 2005; Ometto et al. 2005; Schmid et al. 2005; Heuertz et al. 2006; Li and Stephan 2006; Pyhäjärvi et al. 2007; Ross-Ibarra et al. 2008). However, such simulations still assume unstructured populations, while the empirical data may derive Supporting information is available online at http: / cgi/content/full/genetics /dc1. 1 Corresponding author: Institute of Integrative Biology, Plant Ecological Genetics, ETH Zurich, Universitätstrasse 16, CH-8092 Zurich, Switzerland. thomas.staedler@env.ethz.ch from subdivided species and a variety of sampling schemes. The data typically collected in such projects consist of multilocus DNA sequences from which genealogical information is extracted, e.g., on the number of segregating sites (S), the average pairwise sequence diversity (p), and either the full site frequency spectrum or summary statistics based on the site frequency spectrum, such as Tajima s D (Tajima 1989), Fu and Li s D (Fu and Li 1993), and Fay and Wu s H (Fay and Wu 2000). Here, we are concerned with estimates of specieswide demographic history from such summary statistics and particularly with the joint impact of population subdivision, population expansion, and the sampling scheme on statistics summarizing the site frequency spectrum. Our motivation stems largely from striking patterns in empirical data sets that, we believe, need to be evaluated for their general implications. For example, in empirical studies that include several genuine population samples, it becomes possible to contrast the site frequency spectra of local population samples with those obtained for the pooled (combined) samples. Although this possibility has not been exploited in all such studies, Genetics 182: (May 2009)

2 206 T. Städler et al. the common observation has been that site frequency spectra are more skewed toward low-frequency polymorphisms in pooled samples, compared to the means of genuine population samples; we refer to this phenomenon as the pooling effect. Such genealogical signals have been obtained in a variety of organisms such as humans (Alonso and Armour 2001; Ptak and Przeworski 2002; Hammer et al. 2003), Drosophila (Pool and Aquadro 2006), and several plant species (Ingvarsson 2005; Arunyawat et al. 2007; Moeller et al. 2007; Pyhäjärvi et al. 2007). At least partly due to the putatively confounding effects of population subdivision and population expansion first identified by Ptak and Przeworski (2002), several studies have explicitly favored the analysis and interpretation of data obtained from large samples of single populations or geographic regions (e.g., Adams and Hudson 2004; Marth et al. 2004; Garrigan et al. 2007). Similarly, a recent multilocus study of several populations of teosinte (the wild ancestor of maize) reiterated the notion of genealogical signals for population expansion (e.g., negative values of Tajima s D) being confounded with effects of population structure (Moeller et al. 2007). In our recent study of patterns of polymorphism and population subdivision in two closely related species of wild tomatoes, however, we arrived at quite different conclusions (Arunyawat et al. 2007). Inspired by the results of two complementary simulation studies of subdivided populations (Ray et al. 2003; De and Durrett 2007; see below), we interpreted this apparently widespread empirical pattern (see references cited above) as being a consequence of the effect of the sampling scheme on sample genealogies. Specifically, Arunyawat et al. (2007) envisioned the impact of population subdivision as masking the true genealogical signal of the specieswide demography via loss of low-frequency polymorphisms in local samples, rather than as creating an excess of low-frequency variants in pooled samples. The difference in interpretation compared to the teosinte study by Moeller et al. (2007) is not due to the underlying data, as the site frequency spectra of pooled population samples (four populations per species in our study on wild tomatoes) also showed a marked excess of singletons compared to the average spectra for individual populations, resulting in significantly negative values of Tajima s D and/or Fu and Li s D (Arunyawat et al.2007; Städler et al. 2008). Using parameter values designed to reflect patterns of variation in human mitochondrial DNA, Ray et al. (2003) simulated a range expansion under steppingstone structure and sampled 30 sequences taken from a single deme. Under fairly low levels of gene flow between demes, sample genealogies were shaped by a mixture of local coalescent events (leading to near-zero branch lengths) and ancestral lineages coalescing in the founding deme at the onset (looking forward in time) of the specieswide range expansion. The overall shape of such genealogies yielded values of Tajima s D that were not significant under NE expectations for low to moderate migration rates. Only under very high levels of gene flow did the single-deme samples exhibit genealogies expected for strong population expansions (Ray et al. 2003), reflecting the low number of coalescence events within the sampled deme (Wakeley 1999, 2001, 2004; Wakeley and Aliacar 2001). Analogous phenomena were recently uncovered by De and Durrett (2007), who simulated an equilibrium stepping-stone model of 100 demes, using scaled mutation parameters appropriate for nuclear loci in humans. Compared to expectations for a panmictic equilibrium population, these authors found lower average S, more extended linkage disequilibrium, and average frequency spectra skewed toward intermediate-frequency polymorphisms (i.e., positive Tajima s D values) in simulated samples from single demes. A notable limitation of these studies is their focus on samples obtained from single demes (Ray et al. 2003), although De and Durrett (2007) also considered random samples and showed that their genealogical properties approach those of samples from random-mating populations. Arunyawat et al. (2007) qualitatively interpreted the pooling effect identified in their wild tomato data (and observed for several other empirical studies cited above) as resulting from the changed genealogical structure of combined samples, analogous to increasing the proportion of migrants in single-deme samples. In both circumstances, scattering-phase coalescence events ought to decrease in favor of higher numbers of ancestral lines remaining at the start of the collecting phase (looking backward in time) of the coalescent process, thus yielding higher proportions of external branches of the genealogies. In this study, we set out to test the interpretations of Arunyawat et al. (2007), in particular their qualitative prediction that the genealogies of pooled samples (composed of several local population samples treated as one entity), as reflected in the site frequency spectrum, should be intermediate between those of local and scattered samples. The latter comprise single sequences from each of many demes and are known to possess the genealogical structure of samples from unstructured populations (assuming a large number of demes; Wakeley 1999, 2001, 2004; Wakeley and Aliacar 2001). More generally, we assess the genealogical properties of these three types of samples under both the symmetric finite-island model and the stepping-stone model of population structure. Given that Arunyawat et al. (2007) interpreted the frequent observation of negative Tajima s D values in pooled samples of diverse taxa (e.g., Ptak and Przeworski 2002; Hammer et al. 2003; Pool and Aquadro 2006; Moeller et al. 2007) as signatures of specieswide (range) expansions, our emphasis is on models of structured populations undergo-

3 Genealogies and Population Subdivision 207 ing an overall expansion in population size. To this end, our coalescent simulations encompass a wide range of migration rates, magnitude, and timing of expansions. We find a strong effect of the sampling scheme on two commonly used summary statistics based on the site frequency spectrum, where the pooling effect gradually decreases with increasing levels of gene flow. Our results imply that sequence data for most species should not be evaluated under a panmixia framework and that attention to sampling is paramount even for weakly subdivided taxa. Moreover, disentangling the effects of demography from those of positive selection appears to be an even more challenging task than currently assumed. MODELS AND METHODS Coalescent simulations under two models of population structure: All patterns of sequence diversity were generated using the coalescent simulation software ms (Hudson 2002) to model the following evolutionary scenario. At the time of sampling, the population consists of I demes, each containing N 0 diploid individuals. For the first set of simulations, the subdivided population is at equilibrium with constant population size. Along every line mutations accumulate at rate m, and u ¼ 4N 0 m. We chose to simulate mainly with a fixed value of u ¼ 1 (for simulations implementing 100 demes) rather than with a fixed number of segregating sites S or choosing different u s for different simulation scenarios to obtain realistic values of S and/or p. Simulating with afixeds has been shown to yield inaccurate results under nonequilibrium demography (Ramos-Onsins et al. 2007), and the critical values for Tajima s D and Fu and Li s D depend not too strongly on u. Moreover, the effects we describe in this article are orders of magnitude larger than what could be generated by choosing a different u or fixing S. The demes exchange (haploid) migrants under either an island model or a two-dimensional stepping-stone model at rate m, and we consider a broad range of gene flow: 0.1 # 4N 0 m # 100. Under the island model an ancestral line in deme j switches its location to deme i at rate m/(i 1). Under the stepping-stone model, we assume that I ¼ a 2 ; i.e., the population is arranged in a square lattice of a 3 a demes and we assume periodic boundary conditions. This means that the migration rate is m/4 if i ¼ (i 1, i 2 )andj ¼ (j 1, j 2 ) are neighboring demes and 0 otherwise. Here, i and j are neighbors if j(i 1 j 1 )mod aj 1 j(i 2 j 2 )mod aj ¼1. In other words, an individual at location (a, i 2 ) can migrate to (1, i 2 ) and one at (i 1, a) can migrate to (i 1, 1) and vice versa. We modified ms to be able to efficiently sample sequences from randomly chosen demes rather than fixed demes. Specifically, in each iteration the modified version of ms shuffles the entries in the sample configuration array inconfig at the beginning of the function segtre_mig (Hudson 2002). The C code of this program is available from Implementing range expansions under population subdivision: For an additional set of coalescent simulations, we assume that the structure in the population was created some time t in the past (t is measured in units of 4N 0 generations), i.e., before time t the population was panmictic and of size N A. This scheme ought to be plausible under range expansions, e.g., as exemplified by migration out of Africa by both humans and Drosophila melanogaster and subsequent colonization of expansive areas, or temperate-zone populations expanding from glacial refugia, or following a speciation event such as that inferred for the two wild tomato species Solanum peruvianum and S. chilense (Städler et al. 2005, 2008). Moreover, this is essentially a generalized isolation with migration (IM) model of divergence with a large number of extant demes (Nielsen and Wakeley 2001; Hey and Nielsen 2004; Wilkinson-Herbots 2008). Looking forward in time, at time t the ancestral population splits into I demes of equal size and in equal proportions. The expansion factor for the total population at time t is thus given by b ¼ I 3 N 0 =N A : ð1þ Note that a value of b ¼ 1 implies constant population size in the sense that a panmictic population at time t in the past split into I demes each of size N 0 (¼ N A /I), without changing the total census size of the entire, now subdivided population. This particular scenario can be seen as a form of range fragmentation, albeit without decline in total population size. Sampling schemes and descriptors of diversity and differentiation: Simulated samples of total size n (20 in our numerical examples) from the structured population were implemented as local, pooled, orscattered. Local samples contain n sequences from a single island; i.e., only one arbitrarily chosen deme is sampled. Pooled samples contain several lines each from several demes (we take five lines from each of four demes in our simulations). Scattered samples encompass single sequences from each of n different demes, i.e., only one sequence per sampled deme. A commonly used statistic to quantify population structure from patterns of diversity within and among local populations is F ST (e.g., Hudson et al. 1992). In particular, if the number of demes is large and the population is in equilibrium, E½F ST Š¼1=ð1 1 4N 0 mþ ð2þ (e.g., Wright 1951), and thus migration rates can, in principle, be estimated from observed values of F ST (but see Whitlock and McCauley 1999 for numerous caveats and Jost 2008 for a more fundamental critique of F ST -based estimates of differentiation and gene flow).

4 208 T. Städler et al. In our simulations, F ST can be computed only under the pooled sampling scheme. We used the formula F ST ¼ 1 p w /p b, where p b is the average number of differences for pairs of sequences taken from different demes, and p w is the average number of differences for pairs sampled within demes. The average here means only the average over all pairs of sequences, and not over all simulation runs. Thus, for every simulation run, we recorded exactly one F ST value. For the stepping-stone model, we used the same computations; i.e., F ST does not take isolation by distance into account. To describe sequence diversity patterns, we focus on the site frequency spectrum, as summarized by statistics such as the widely used Tajima s D (Tajima 1989) and Fu and Li s D (Fu and Li 1993). For clarity, we denote these distinct D statistics as D T and D FL, respectively. Using these particular summary statistics enables us to perform power analyses to reject the standard NE model. Moreover, we chose to include D FL because the singleton class appeared to be the major reason for the lower/ more negative D T values in pooled vs. local samples of wild tomatoes (Arunyawat et al. 2007). The statistical package R was used to drive our modified version of ms and to compute these statistics from its output; the corresponding R scripts are available at evolbio.mpg.de/sampling. Demographic signatures in multilocus data from wild tomatoes: Our recent studies on population subdivision and speciation in two wild tomato species provided the motivation for our current simulation study (Roselius et al. 2005; Städler et al. 2005, 2008; Arunyawat et al. 2007; see Introduction). We extracted the biallelic, synonymous and noncoding single-nucleotide polymorphisms (SNPs) contained in the data of Arunyawat et al. (2007) to confine analyses of the site frequency spectrum to putatively neutral polymorphisms and those fitting the infinite-sites assumption implemented in Hudson s (2002) ms program. Tajima s D T and Fu and Li s D FL values of these pruned data sets were computed in the SITES program ( Specifically, we computed the average D T and D FL values across the four single-population samples per species, as well as the corresponding statistics obtained by treating all sequences within species as a single sample; we refer to these latter entities as the pooled samples (see Arunyawat et al. 2007). While devising estimators of the demographic parameters b and t is beyond the scope of the present article (and may require the availability of multilocus data obtained from scattered samples), we implemented coalescent simulations reflecting the actual sampling scheme used by Arunyawat et al. (2007) and fixed other parameters according to aspects of the empirical data and perceived biological context. Both stepping-stone and island-model simulations were performed with 400 local demes, u ¼ 0.25, and sampling 11 sequences from each of 4 local demes (recall that demes were randomly chosen for each of the 1000 iterations). The migration rate was set to 4N 0 m ¼ 5, reflecting the average F ST estimates found for both species (Arunyawat et al. 2007; and see Figure 5 below). Simulations explored the joint effects of varying the time (t) and magnitude (b) of demographic expansions concomitant with establishment of population subdivision. From these simulations, we recorded estimates of D T and D FL for two sampling schemes, local (means of the 4 local samples) and pooled (all 44 sequences treated as one sample). Average estimates of D T and D FL obtained in this manner for many simulated combinations of b and t are presented as contour plots. RESULTS Our coalescent simulations yield results for levels of nucleotide diversity, summary statistics based on the site frequency spectrum, and population differentiation under both equilibrium and nonequilibrium demographic history, all obtained for three different sampling schemes: local, pooled, and scattered. All our findings are consistent between the island model and the stepping-stone model, and in this article we focus on quantitative results obtained under stepping-stone spatial structure; results for the island model are presented as supporting figures online. The site frequency spectrum under population subdivision: The simplest demographic scenario we analyzed is an equilibrium population subdivided into 100 demes; i.e., we first focus on the effects of population structure per se without any past changes of population size. As expected, characteristics of the site frequency spectra depend strongly on the sampling scheme. For the stepping-stone model of population structure, Figure 1 shows the simulation results under various levels of gene flow. For a migration rate of 4N 0 m ¼ 10, local samples produce values of Tajima s D T (Fu and Li s D FL )thatare significantly different from values expected under the NE model (two-tailed test, P, 0.05) in 16% (39%) of all cases, while scattered samples give significant results in only 6% (7%) of all simulations. In particular, we see that local samples generate values for both statistics that are higher than expected for samples from panmictic populations, reflecting a site frequency distribution skewed toward intermediate-frequency mutations; this result mirrors the recent work of De and Durrett (2007). For migration rates (in units of 4N 0 m) between 2 and 50, pooled samples exhibit site frequency spectra that are broadly intermediate between those of local and scattered samples. The differences in sample genealogies (as reflected in estimates of D T and D FL ) gradually diminish with higher levels of gene flow, but some differences among sampling schemes are still apparent at fairly high migration rates (e.g., 4N 0 m. 50 for D T and 4N 0 m 100 for D FL ; Figure 1). Importantly, pooling data from several subpopulations does not generate negative

5 Genealogies and Population Subdivision 209 Figure 1. Averages of Tajima s D T (A) and Fu and Li s D FL (B) under three sampling schemes as a function of levels of gene flow. We simulated an equilibrium stepping-stone model with I ¼ 100 islands; the simulations were carried out without recombination. Every plotted point is based on 1000 independently generated data sets, and standard errors of the means are indicated by vertical lines. values of D T or D FL without an expansion of the total population (see discussion). These observations also hold for the island model, albeit with smaller discrepancies between the summary statistics for scattered samples and those for the two other sampling schemes (i.e., a less pronounced skew toward intermediatefrequency mutations for both pooled and local samples; supporting information, Figure S1). For lower levels of gene flow (4N 0 m, 2), the site frequency spectra of local samples gradually shift toward NE expectations, while those of pooled samples yield increasingly positive values of both D T and D FL under decreasing levels of migration. We checked our simulations of this equilibrium model against analytical results showing that local samples ought to be invariant for the level of nucleotide diversity, p, irrespective of the level of symmetrical interdeme migration in an island model (Slatkin 1987; Strobeck 1987). For both models of population structure, we found approximately invariant mean p-values for local samples over the entire range of simulated migration rates (results not shown). The impact of population/range expansions on the site frequency spectrum: Next, we considered scenarios of (range) expansions under concomitant establishment of subdivision of the total species range, as described in the models and methods section. In particular, we simulated a single ancestral population that at time t before the present experienced a fragmentation into I demes, where the total number of individuals across the subdivided population could vary from N A (the number of individuals in the single ancestral population, in which case the expansion factor b ¼ 1) to N A (equivalent to b ¼ 100). Under a steppingstone model with a 10-fold population expansion 8N 0 generations in the past, local samples still exhibit values of D T and D FL that would be expected under NE conditions, as long as gene flow is fairly low (Figure 2); these findings are consistent with simulation results by Ray et al. (2003). The corresponding simulation results for the island model are shown in Figure S2. Scattered samples, however, contain a clear signal of the specieswide expansion at any level of gene flow. The reason is that the coalescent of scattered samples almost behaves like a neutral one with a population size proportional to the number of demes. Hence, this coalescent picks up the signal of expansion as in the panmictic case. The same should hold for any sampling scheme under sufficiently high migration, but Figure 2 shows that even under high levels of gene flow (4N 0 m ¼ 100), the site frequency spectrum of local samples is still different from that of pooled and scattered samples. Moreover, the simulation results plotted in Figure 2 were obtained under a 10-fold expansion, and we may expect discrepancies between local and scattered samples to extend to even higher migration rates with higher expansion factors (see Figure 3). Again, D T and D FL values obtained for pooled samples are intermediate between those of local and scattered samples except Figure 2. Averages of Tajima s D T (A) and Fu and Li s D FL (B) under specieswide expansion (b ¼ 10, t ¼ 2) as a function of migration rates between demes. We simulated a stepping-stone model with 100 demes without recombination. Every plotted point is based on 1000 independently generated data sets; error bars represent standard errors.

6 210 T. Städler et al. Figure 3. Averages of Tajima s D T (A) and Fu and Li s D FL (B) as functions of the expansion factor b (with fixed migration rate and time of expansion), with the power of these statistics evaluated in the top part of each plot (right y-axis; power was assessed at a level of P ¼ 0.05). The same simulation scheme as in Figure 2 was used. for very low migration rates (4N 0 m, 0.5), similar to the case of equilibrium subdivided populations (Figures 1 and 2). These simulation results imply that both local and pooled samples may be expected to underestimate the extent of any specieswide (range) expansion to various degrees, depending on levels of gene flow connecting the demes and the age and magnitude of the expansion. The latter aspect, however, is influenced by our choice of simulating an instantaneous expansion followed by a period of constant population size until the time of sampling; an exponential expansion scheme until the present would yield higher detectability of expansion from local and pooled samples. We point out that quantitative details of our simulation results for the three sampling schemes depend to some extent on our choice of sampling 20 sequences distributed over one ( local ), four (pooled), and 20 demes ( scattered ), respectively. Generally speaking, decreasing the number of sequences sampled per deme and/or increasing the number of demes that sequences are sampled from shifts the site frequency spectrum more toward that of scattered samples, reflecting the diminished impact of the scattering phase of the coalescent process on such samples. The exact numerical composition of local and pooled samples also affects the migration rate at which the expected D T and D FL values for pooled samples drop below those for local samples (see Figures 1 and 2). Power of D T and D FL to detect expansion under different sampling schemes and expansion times: Illustrated by an intermediate level of interdeme migration (4N 0 m ¼ 10), we assessed the power of the test statistics D T and D FL under a range of expansion factors and times of expansion. For the three sampling schemes, Figure 3 summarizes the power of D T and D FL assuming an expansion time of t ¼ 8N 0 generations ago. The low power of local samples to detect significant departures from the NE model regardless of the magnitude of the specieswide expansion is striking; qualitatively, this is consistent with results by Ray et al. (2003). Even if the specieswide expansion was 100-fold, local samples deviate from NE expectations in only 8% (for D T ) and 12% (for D FL ) of all cases under these conditions. In sharp contrast, scattered samples deviate from NE expectations in 98% (99%) of all cases for b ¼ 100. The corresponding simulation results for the island model are presented in Figure S3. Next, we illustrate the effect of the timing of the expansion for a fixed 10-fold expansion and a migration rate of 4N 0 m ¼ 10. Under these conditions, expansion times in the approximate range 1, t, 15 can be detected in principle, but again with striking differences in power exhibited by local vs. scattered samples (Figure 4). The shape of the curves in Figure 4 can be explained intuitively: if t is very small (i.e., establishment of population structure was very recent), samples appear to be drawn from a panmictic population of constant size. On the other hand, if t is very large, samples appear to be drawn from an equilibrium subdivided population and hence the expansion cannot be detected. Under the Figure 4. The power of Tajima s D T (A) and Fu and Li s D FL (B) as a function of the expansion time t (with fixed migration rate and expansion factor). The same simulation scheme as in Figure 2 was used; power was assessed at a level of P ¼ 0.05.

7 Genealogies and Population Subdivision 211 Figure 5. F ST as a function of 4N 0 m. (A) t ¼ 8N 0 generations is fixed and b is varying; (B) b ¼ 10 is fixed and t is varying. The same simulations as in Figure 2 were used (i.e., stepping-stone model with 100 demes). island model, all results are qualitatively the same but with even smaller power to reject stable population size for local samples (at most 10% less power; Figure S4). As these power assessments are based on simulated sample genealogies of single loci, the actual power available with empirical multilocus data would be correspondingly higher. All simulation results appear to depend only weakly on the number of demes, as long as I? n. However, an increase in the number of demes carries important connotations, as the genealogical signatures of past expansions remain detectable for longer time periods than with fewer demes. For example, simulating with a constant ratio t/i ¼ 0.02 (i.e., equivalent to t ¼ 2 for I ¼ 100, t ¼ 10 for I ¼ 500, etc.) under otherwise equal demographic conditions, we obtained fairly constant but sampling-specific estimates of D T and D FL, as illustrated for the island model in Figure S5. One interpretation is that the effective population size is proportional to I but depends on the sampling scheme. These observations imply approximately equal detectability of expansions (for a given sampling scheme) over a large range of I and thus the dependency of relevant t-values on I. This latter effect is a direct consequence of lengthening the collecting phase of the coalescent process with increasing numbers of demes in the total population. Estimates of F ST under nonequilibrium conditions: We studied the behavior of F ST in expanding populations (cf. models and methods) and consider both the stepping-stone and the island model of population structure. Given such scenarios, estimates of F ST depend strongly on the migration rate 4N 0 m but only weakly on the expansion time t and the expansion factor b, as long as 4N 0 m $ 2 (Figure 5 and Figure S6). Under moderate and high levels of gene flow, Equation 2 gives a reasonable approximation of the simulated results, although the level of population structure is mostly higher than expected under equilibrium conditions, and thus actual levels of gene flow would be mostly underestimated using F ST calculated from nonequilibrium populations. Under low levels of gene flow (4N 0 m, 2), however, strong and/or recent expansions yield less population differentiation than expected under the assumption of equilibrium conditions, and thus levels of gene flow would be mostly overestimated (Figure 5; see also Excoffier 2004). Reanalyzing the pooling effect in two species of wild tomatoes: Our reanalysis of the multilocus wild tomato data, limited to silent sites and biallelic SNPs, confirms the patterns originally identified by Arunyawat et al. (2007) based on all SNPs; with the exception of one locus in S. peruvianum (CT066, D FL ) and one locus in S. chilense (CT268, D T ), the site frequency spectra at individual loci yield lower/more negative values of D T and D FL in the pooled samples compared to the means of four samples representing local populations (Table 1). Considering the pooled samples, multilocus average D T (D FL ) values are 1.14 ( 1.38) in S. peruvianum and 0.30 ( 0.76) in S. chilense; the magnitude of the drop from single-population means in D T (D FL ) is 0.89 (1.04) in S. peruvianum and 0.78 (1.31) in S. chilense (Table 1). Given our simulation results under both equilibrium and expansion scenarios, the observed summary statistics for pooled samples of both species are incompatible with equilibrium assumptions, but compatible with specieswide expansions (Figures 1 3). On the basis of the observed levels of population subdivision (F ST ; Arunyawat et al. 2007), a rough estimate of average migration rates is 4N 0 m 5 (Figure 5). Our simulation results presented above convey how summary statistics such as D T and D FL obtained from local or pooled samples ought to be affected jointly by migration rates and both the magnitude and the time of expansion. Figure 6 presents results of coalescent simulations tailored to fit the empirical wild tomato data (i.e., sampling 11 sequences from each of four local demes and fixing the migration rate to 4N 0 m ¼ 5). Even when restricting the analysis to two parameters that were allowed to vary widely, it is clear that given observed values of D T and D FL are not sufficient to permit robust estimation of the demographic parameters t and b. The simulated average D T values for pooled samples shown in the contour plot (Figure 6B) suggest that the observed pooled D T values are compatible with $25-fold expansion in S. peruvianum and with $4-fold expansion in S. chilense.

8 212 T. Städler et al. Locus TABLE 1 Tajima s D and Fu and Li s D values in pooled vs. population-specific tomato samples S. peruvianum S. chilense Mean of four populations Pooled Mean of four populations Pooled CT CT CT CT CT CT CT CT Average For each locus, the first row shows values of Tajima s D T, while the second row shows values of Fu and Li s D FL. Data were taken from Arunyawat et al. (2007), but here all nonsynonymous SNPs and those segregating more than two nucleotides (i.e., showing evidence of recurrent mutations) were eliminated before estimating D T and D FL. The pooled columns show values based on the combined (total) sample within each species, treated as a single entity. Note that almost all loci individually exhibit lower/more negative D T and D FL values in the pooled samples. Locus CT208 was not included for S. chilense because patterns of polymorphism were suggestive of natural selection (Arunyawat et al. 2007). When viewed in isolation, the observed local D T values are consistent with a very high expansion factor for S. peruvianum (b. 200) and with b 1 for S. chilense. However, considering the observed differences local vs. pooled,.40-fold and,4-fold expansions appear to be inconsistent with the data for both species (Figure 6; Table 1). Similarly, the observed D FL values would seem to be compatible with.150-fold expansion in S. peruvianum and with.3-fold expansion in S. chilense, whereas on the basis of the observed differences local vs. pooled, only.120-fold expansions appear to be compatible with the data for both species (Table 1; Figure S7). Assuming that the expansion time of both species mirrors their speciation time, we note that these b- values are larger than those assuming the model of Wakeley and Hey (1997) as obtained in Städler et al. (2008). This apparent discrepancy can be explained at least partly by our findings, since the Wakeley Hey model treats any empirical sample as one from a panmictic population and hence picks up signals of expansion only partly, leading to smaller estimates of the expansion factor. We emphasize that changing some of the simulated parameter values, such as the number of demes and the migration rate among demes, would suggest other b-values as being consistent with the empirical data. Similarly, assuming an island model rather than a stepping-stone structure would result in a site frequency spectrum more biased toward low-frequency variants, hence implying signatures of more moderate expansions than suggested by some features of the data mentioned above (Figure S8, Figure S9). Moreover, these simulations were performed with a uniform migration rate across the entire subdivided population, whereas the empirical data may reflect (perhaps dramatic) differences in immigration rates among local samples. Irrespective of the inherent difficulties of inferring credible demographic parameters under population subdivision and nonstationarity, the generally lower/ more negative D T and D FL values for both local and pooled S. peruvianum samples strongly suggest a more pronounced expansion compared to S. chilense (Table 1; Arunyawat et al. 2007; Städler et al. 2008). DISCUSSION Although models of subdivided populations have been studied extensively, the lack of analytical results for local samples of size.2 may have restricted their utility for genealogical inference. Wilkinson-Herbots (2008) recently obtained expressions for the expected mean and variance of pairwise sequence divergence (samples of size 2 drawn either from the same subpopulation or from different subpopulations in a finite island model) in a generalized IM model (see also Excoffier 2004). As our simulations were motivated mainly by patterns in empirical data, we have focused on larger sample sizes and aspects other than pairwise sequence diversity, but the underlying model of population subdivision with variable timing and magnitude of population expansion is identical to that studied by Wilkinson-Herbots (2008). For the most part, our simulations implemented 100 demes and a sample size of 20, and thus should approach the many-demes limit utilized in Wakeley s studies of the island model (e.g., Wakeley 1999, 2001, 2004; Wakeley and Aliacar 2001) and more recently shown by De and Durrett (2007) for the stepping-stone model. Simulating local samples under both island and stepping-stone spatial structure, De and Durrett (2007) found lower numbers of segregating sites, higher linkage disequilibrium (LD), and median site frequency spectra that were skewed toward intermediate-frequency polymorphisms (i.e., yielding positive D T values) in comparison with samples from randomly mating populations. However, they did not investigate the behavior of samples pooled across several demes. Wakeley and Lessard (2003) have shown that levels of LD can be significantly elevated in local samples, depending on migration rates

9 Genealogies and Population Subdivision 213 Figure 6. Contour plots showing average values of D T obtained in simulations designed to mimic the empirical sampling of Arunyawat et al. (2007). (A) Single-deme samples (n ¼ 11 sequences); (B) pooled samples (n ¼ 44 sequences). For each of the 1000 simulations for a given combination of expansion factor (b) and time since expansion (t), 11 sequences were drawn from each of 4 randomly chosen demes, and the 44 sequences were also evaluated as a pooled sample (B). We simulated 400 demes under a stepping-stone structure [u (per deme) ¼ 0.25, 4N 0 m ¼ 5]. The empirical average D T values for S. peruvianum are 0.25 (local) and 1.14 (pooled), while they are 0.48 (local) and 0.30 (pooled) for S. chilense ; see Table 1. and the total number of demes. Their analyses highlighted the potential impact of sampling scheme and demography on the validity of evolutionary inferences drawn under unrealistic assumptions, a topic to which we shall return below. Site frequency spectra under different sampling schemes: Using coalescent simulations, we evaluated three sampling schemes of sequences drawn from subdivided populations under different levels of migration among local demes; all three sampling schemes have corresponding examples in empirical studies of DNA sequence diversity designed to infer aspects of demographic history and natural selection from the site frequency spectrum and/or patterns of LD. Our simulations of subdivided populations of constant total size (Figure 1) have identified a wide range of migration rates where the pooling effect, as interpreted by Arunyawat et al. (2007), appears to hold: local samples are characterized by higher values of Tajima s D T and Fu and Li s D FL than pooled samples for Nm (immigrants per generation). 0.5, and the underlying genealogies of such samples converge in shape only at very high levels of gene flow (Nm. 25). With low migration levels (Nm, 0.5), the pattern is reversed in that pooled samples show an excess of highly positive D T and D FL values. At these low gene flow levels, local samples are expected to undergo a series of coalescent events within the sampled demes before entering the collecting phase with a single ancestral line per deme. The resulting long internal branches of genealogies encompassing lineages from several local demes (i.e., pooled samples) typically exhibit multiple fixed mutations between different demes, leading to intermediate-frequency SNPs in pooled samples. Pannell (2003; his Table 3) has addressed these effects of local vs. pooled samples in the context of metapopulation dynamics. That these genealogical effects are mostly in the opposite direction, i.e., that pooling several local samples is expected to increase the proportion of lowfrequency polymorphisms under intermediate and high levels of migration, has until recently been obscure, even though simulation results by Pannell (2003; his Table 3) indicated such an effect. Analyzing previously published human data sets encompassing a variety of sampling schemes, Ptak and Przeworski (2002) identified a pattern of increasing skew toward low-frequency mutations (i.e., negative D T ) with more geographically heterogeneous sampling (equivalent to pooling across demes) and concluded that signatures of population expansion may be confounded with effects of population subdivision. Our coalescent simulations of equilibrium island/stepping-stone models under a reasonable range of migration rates show that pooling does not lead to negative D T values and thus do not support this particular interpretation of a confounding effect of subdivision. On the contrary, pooled samples (albeit to a lesser extent than local samples) still exhibit the signatures of local (scattering-phase) coalescent events (i.e., lower proportion of external branch length of the sample genealogies), compared to scattered samples whose genealogies should be roughly equivalent in structure to those from panmictic populations (but see below). This intermediate frequency spectrum of pooled samples (between the extremes of local and scattered samples, respectively) is what Arunyawat et al. (2007) predicted on the basis of their genealogical interpretation of the pooling effect. For the case of nonequilibrium populations, one caveat that ought to be mentioned is that signatures of expansion will be seen in scattered samples under low levels of gene flow even without any increase in the total number of individuals (i.e., b#1). The reason is that at time t, the establishment of subdivision increases the effective population size by a factor of 1 1 (1/4N 0 m)(wakeley 2000). Consequently, an apparent expansion may be seen under strong subdivision (e.g., 4N 0 m, 2); the increasingly negative values for the scattered samples shown in Figure 2 with decreasing migration rate appear to reflect this phenomenon. Site frequency spectra and sampling in nonequilibrium subdivided populations: We found effects of the sampling scheme on sample genealogies up to very high migration rates, especially under strong specieswide expansions (Figures 2 and 3). In other words, even

10 214 T. Städler et al. under high migration rates (e.g., Nm ¼ 25), local samples may not adequately reflect the specieswide demography. The relevance of our simulation findings is essentially an empirical issue that should be judged by its predictive power and relevant data. Our simulations predict a pooling effect over a wide range of migration rates, implying one way to test the null hypothesis of panmixia, i.e., that the genealogies of local samples are indistinguishable from those of scattered samples. Very few published studies have performed separate analyses of local population samples and the pooled sample consisting of several local samples, but many studies can, in principle, be used to test our predictions because several local samples have been included [Drosophila (Baudry et al. 2004, 2006; Nolte and Schlötterer 2008), humans (Marth et al. 2004; Voight et al. 2005; Keinan et al. 2007; Garrigan et al. 2007), and plants (Heuertz et al. 2006; Pyhäjärvi et al. 2007)]. Pool and Aquadro (2006) demonstrated the pooling effect in sub-saharan D. melanogaster, as have several studies in plants (Ingvarsson 2005; Arunyawat et al. 2007; Moeller et al. 2007) and humans (e.g., Ptak and Przeworski 2002; Hammer et al. 2003). In all these cases, the site frequency spectra of pooled samples were characterized by an excess of low-frequency variants, resulting in (more) negative D T values. Some of these studies explicitly considered the possible contribution of purifying selection to these patterns, but concluded that demographic (or range) expansions were at least partly responsible for the skew in the site frequency spectra (Pool and Aquadro 2006; Moeller et al. 2007; see also Nordborg et al. 2005). As far as neutral polymorphisms are concerned, our simulation results fully concur with these interpretations, as only global expansions and not the effects of population subdivision per se can generate substantially negative D T values (Figures 1 3). As many other researchers before us, we have investigated the properties of general models of population structure in the context of sampling schemes and their effects on sample genealogies. In contrast, Hammer et al. (2003) presented coalescent simulations of a nonstandard continuous-splitting model of human population structure (where demes do not exchange migrants after being isolated from other demes) and concluded that site frequency spectra of human data sets may be explained without invoking global expansion. Levels of gene flow and within-species panmixia: It may seem surprising to observe the pooling effect even in highly mobile species such as D. melanogaster (Pool and Aquadro 2006), thus providing clear-cut evidence for deviations from panmixia at seemingly low levels of population subdivision (e.g., as quantified by F ST ). The impression of low levels of population substructure may, at least in part, be a consequence of the F ST estimator most widely used for DNA sequence data, F ST ¼ 1 p w /p b, which treats each SNP as a separate locus (Hudson et al. 1992). In empirical as well as simulated data sets, many segregating sites that appear as singletons in local population samples remain singletons in pooled samples. These sites do not contribute much to nucleotide diversity in either local or total samples, compared to SNPs with intermediate frequencies, but they fuel a more pronounced increase in the number of segregating sites in pooled samples [i.e., Watterson s (1975) estimator u W increases much more than p; seealsoray et al. 2003]. These underlying features explain the effect of pooling under apparently low levels of population subdivision. Moreover, they also highlight the inherent dangers of equating low F ST estimates with the (near) absence of population subdivision. In the framework of the infinite-alleles model, Jost (2008) recently argued that F ST (and G ST ) are not at all appropriate measures of differentiation and pinpointed mathematical misconceptions underlying the standard approach. In particular, he identified the classical additive partitioning of total heterozygosity (H T,the heterozygosity of the pooled subpopulations) into mean within-subpopulation heterozygosity (H S ) and a between-subpopulation component (H T H S ¼ D ST ; Nei 1973) as erroneous, because H S and D ST are not truly independent but rather related through an incomplete partitioning ( Jost 2008). Because classical results concerning the absolute number of migrants per generation required to prevent significant differentiation among local demes in the finite-island and steppingstone models (e.g., Nm. 1, Equation 2 above; e.g., Wright 1951; Maruyama 1971) are based on the interpretation of F ST as a measure of differentiation, Jost (2008) concluded that such rules of thumb must be considered invalid. Although Jost (2008) did not formally evaluate the infinite-sites model of sequence evolution, this conclusion goes hand in hand with our interpretations in the preceding paragraph. Importantly, the marked effects of sampling scheme on the site frequency spectrum in our simulations despite high levels of gene flow are not restricted to expanding populations but were found also for subdivided populations at demographic equilibrium (e.g., Nm. 10; Figures 1 and 2). Similar results were previously obtained by De and Durrett (2007) for equilibrium populations and, implicitly, by Ray et al. (2003) for expanding populations. Empirical considerations and implications for statistical inference: Large-scale population-genetic data sets are commonly evaluated in the framework of panmictic populations undergoing temporal changes in population size [humans (e.g., Marth et al. 2004; Voight et al. 2005; Garrigan et al. 2007), Drosophila (e.g., Ometto et al. 2005; Li and Stephan 2006; Thornton and Andolfatto 2006), and plants (Heuertz et al. 2006; Pyhäjärvi et al. 2007)]. In conjunction with the empirical data discussed above, our results suggest that a more appropriate approach would acknowledge the subdivided nature of these species explicitly, which

11 Genealogies and Population Subdivision 215 would require paying particular attention to sampling schemes and their genealogical consequences. Especially if empirical sampling is from a local population (as has often been the case), it is inappropriate to compare patterns of diversity with expectations under the classical NE conditions embodied in commonly used tests of neutrality. Generally, local population samples should not be modeled as coming from a panmictic (e.g., continental) population but rather as being drawn from one of many demes composing the total metapopulation. Put another way, the traditional almost exclusive focus on the temporal trajectory of population-size changes that best explains properties of local/regional samples ought to be questioned, especially in systems with moderate-to-high levels of gene flow where sample genealogies are not regionally monophyletic, but rather are deeply embedded in specieswide sample genealogies (Wakeley 2001; Wakeley and Aliacar 2001). In such cases, the properties of local samples (e.g., the number of segregating sites, nucleotide diversity, the site frequency spectrum, and the extent and decay of LD) arguably reflect the level of connectivity with (degree of isolation from) other demes during the scattering phase of the coalescent process (Wakeley 2001; Ray et al. 2003; Wakeley and Lessard 2003; De and Durrett 2007). Our simulation scheme has been simplistic in assuming equal migration rates and deme sizes through time and across the entire metapopulation, but it is straightforward to explain empirical patterns, such as those found for various human population samples, in terms of differences in local population size and immigration rates during the recent past (Wakeley 2001; Ray et al. 2003; Excoffier 2004), arguably mediated by environmental heterogeneity (Wegmann et al. 2006). Guided by notions of the special importance of population structure and the potential for local adaptation in plants, several recent studies have emphasized the need to include genuine local population samples in population-genetic studies, rather than relying exclusively on scattered samples (Wright and Gaut 2005; Moeller et al. 2007; Ross-Ibarra et al. 2008). While we agree that sampling locally has particular merit for studying signatures of local adaptation (see Arunyawat et al. 2007), our present results caution against analyzing such local samples naively, i.e., in the conventional framework of panmictic populations. In principle, sampling several local demes offers the opportunity to empirically contrast the properties of local samples (e.g., site frequency spectrum, decay of LD) with those of the pooled sample, analogous to two of our three simulation sampling schemes. At the very least, this exercise has the potential to infer general features of the specieswide demographic history despite the biases inherent in local samples. It would also (at least partially) mitigate against inferring spurious signatures of natural selection from levels and patterns of nucleotide polymorphism or from the extent of haplotype structure (for examples, see Wakeley and Lessard 2003; De and Durrett 2007). However, only the genealogical structure of scattered samples is comparable to that of an elusive panmictic population that has experienced the same temporal demographic history (e.g., expansion), whereas such samples preclude finding molecular evidence of local adaptation. We thank Tanja Pyhäjärvi for information on her multideme data set on Pinus sylvestris and Aurelien Tellier for valuable discussions on metapopulation dynamics and the coalescent in subdivided populations. Two anonymous referees provided constructive criticism that helped to improve the final version of this article. Furthermore, we thank Dick Hudson for allowing us to redistribute a modified version of his coalescent simulation program ms. This work was funded by the Deutsche Forschungsgemeinschaft through its Priority Program Radiations Origins of Biological Diversity (SPP-1127), grant Ste 325/5-3 (to W.S.), and through Research Unit Natural Selection in Structured Populations (FOR-1078), grant Ste 325/12-1 (to W.S.) and grant Pf 672/1-1 (to P.P.). P.P. acknowledges additional support by the German Federal Ministry of Education and Research through the Freiburg Initiative for Systems Biology, grant LITERATURE CITED Adams, A. M., and R. R. Hudson, 2004 Maximum-likelihood estimation of demographic parameters using the frequency spectrum of unlinked single-nucleotide polymorphisms. Genetics 168: Alonso, S., and J. A. L. Armour, 2001 A highly variable segment of human subterminal 16p reveals a history of population growth for modern humans outside Africa. Proc. Natl. Acad. Sci. USA 98: Arunyawat, U., W. Stephan and T. Städler, 2007 Using multilocus sequence data to assess population structure, natural selection, and linkage disequilibrium in wild tomatoes. Mol. Biol. Evol. 24: Baudry, E., B. Viginier and M. Veuille, 2004 Non-African populations of Drosophila melanogaster have a unique origin. Mol. Biol. Evol. 21: Baudry, E., N. Derome,M.Huet and M. Veuille, 2006 Contrasted polymorphism patterns in a large sample of populations from the evolutionary genetics model Drosophila simulans. Genetics 173: De, A., and R. Durrett, 2007 Stepping-stone spatial structure causes slow decay of linkage disequilibrium and shifts the site frequency spectrum. Genetics 176: Excoffier, L., 2004 Patterns of DNA sequence diversity and genetic structure after a range expansion: lessons from the infinite-island model. Mol. Ecol. 13: Fay, J. C., and C.-I Wu, 2000 Hitchhiking under positive Darwinian selection. Genetics 155: Fu, Y.-X., and W.-H. Li, 1993 Statistical tests of neutrality of mutations. Genetics 133: Garrigan, D., S. B. Kingan, M. M. Pilkington, J. A. Wilder, M. P. Cox et al., 2007 Inferring human population sizes, divergence times and rates of gene flow from mitochondrial, X and Y chromosome resequencing data. Genetics 177: Hammer, M. F., F. Blackmer, D. Garrigan, M. W. Nachman and J. A. Wilder, 2003 Human population structure and its effects on sampling Y chromosome sequence variation. Genetics 164: Heuertz, M., E. De Paoli, T. Källman, H. Larsson, I. Jurman et al., 2006 Multilocus patterns of nucleotide diversity, linkage disequilibrium and demographic history of Norway Spruce [Picea abies (L.) Karst]. Genetics 174: Hey, J., and R. Nielsen, 2004 Multilocus methods for estimating population sizes, migration rates and divergence time, with applications to the divergence of Drosophila pseudoobscura and D. persimilis. Genetics 167:

12 216 T. Städler et al. Hudson, R. R., 2002 Generating samples under a Wright-Fisher neutral model of genetic variation. Bioinformatics 18: Hudson, R. R., M. Slatkin and W. P. Maddison, 1992 Estimation of levels of gene flow from DNA sequence data. Genetics 132: Ingvarsson, P. K., 2005 Nucleotide polymorphism and linkage disequilibrium within and among natural populations of European aspen (Populus tremula L., Salicaceae). Genetics 169: Jost, L., 2008 G ST and its relatives do not measure differentiation. Mol. Ecol. 17: Keinan, A., J. C. Mullikin,N.Patterson and D. Reich, 2007 Measurement of the human allele frequency spectrum demonstrates greater genetic drift in East Asians than in Europeans. Nat. Genet. 39: Li, H., and W. Stephan, 2006 Inferring the demographic history and rate of adaptive substitution in Drosophila. PLoS Genet. 2: e166. Marth, G. T., E. Czabarka, J. Murvai and S. T. Sherry, 2004 The allele frequency spectrum in genome-wide human variation data reveals signals of differential demographic history in three large world populations. Genetics 166: Maruyama, T., 1971 Analysis of population structure II. Two-dimensional stepping stone models of finite length and other geographically structuredpopulations.ann.hum.genet.35: Moeller, D. A., M. I. Tenaillon and P. Tiffin, 2007 Population structure and its effects on patterns of nucleotide polymorphism in teosinte (Zea mays ssp. parviglumis). Genetics 176: Nei, M., 1973 Analysis of gene diversity in subdivided populations. Proc. Natl. Acad. Sci. USA 70: Nielsen, R., andj. Wakeley, 2001 Distinguishing migration from isolation: a Markov chain Monte Carlo approach. Genetics 158: Nolte, V., and C. Schlötterer, 2008 African Drosophila melanogaster and D. simulans populations have similar levels of sequence variability, suggesting comparable effective population sizes. Genetics 178: Nordborg, M., T. T. Hu, Y. Ishino, J. Jhaveri, C. Toomajian et al., 2005 The pattern of polymorphism in Arabidopsis thaliana. PLoS Biol. 3: e196. Ometto, L., S. Glinka, D. De Lorenzo and W. Stephan, 2005 Inferring the effects of demography and selection on Drosophila melanogaster populations from a chromosome-wide scan of DNA variation. Mol. Biol. Evol. 22: Pannell, J. R., 2003 Coalescence in a metapopulation with recurrent local extinction and recolonization. Evolution 57: Pool, J. E., and C. F. Aquadro, 2006 History and structure of sub- Saharan populations of Drosophila melanogaster. Genetics 174: Ptak, S. E., and M. Przeworski, 2002 Evidence for population growth in humans is confounded by fine-scale population structure. Trends Genet. 18: Pyhäjärvi, T., M. R. García-Gil, T. Knürr, M. Mikkonen, W. Wachowiak et al., 2007 Demographic history has influenced nucleotide diversity in European Pinus sylvestris populations. Genetics 177: Ramos-Onsins, S. E., S. Mousset, T. Mitchell-Olds and W. Stephan, 2007 Population genetic inference using a fixed number of segregating sites: a reassessment. Genet. Res. 89: Ray, N., M. Currat and L. Excoffier, 2003 Intra-deme molecular diversity in spatially expanding populations. Mol. Biol. Evol. 20: Roselius, K., W. Stephan and T. Städler, 2005 The relationship of nucleotide polymorphism, recombination rate and selection in wild tomato species. Genetics 171: Ross-Ibarra, J., S. I. Wright, J. P. Foxe, A. Kawabe, L. DeRose- Wilson et al., 2008 Patterns of polymorphism and demographic history in natural populations of Arabidopsis lyrata. PLoS One 3: e2411. Schmid, K. J., S. E. Ramos-Onsins, H. Ringys-Beckstein, B. Weisshaar and T. Mitchell-Olds, 2005 A multilocus sequence survey in Arabidopsis thaliana reveals a genome-wide departure from a neutral model of DNA sequence polymorphism. Genetics 169: Slatkin, M., 1987 The average number of sites separating DNA sequences drawn from a subdivided population. Theor. Popul. Biol. 32: Städler, T., K. Roselius and W. Stephan, 2005 Genealogical footprints of speciation processes in wild tomatoes: demography and evidence for historical gene flow. Evolution 59: Städler, T., U. Arunyawat and W. Stephan, 2008 Population genetics of speciation in two closely related wild tomatoes (Solanum section Lycopersicon). Genetics 178: Strobeck, C., 1987 Average number of nucleotide differences in a sample from a single subpopulation: a test for population subdivision. Genetics 117: Tajima, F., 1989 Statistical method for testing the neutral mutation hypothesis by DNA polymorphism. Genetics 123: Thornton, K., and P. Andolfatto, 2006 Approximate Bayesian inference reveals evidence for a recent, severe bottleneck in a Netherlands population of Drosophila melanogaster. Genetics 172: Voight, B. F., A. M. Adams, L.A.Frisse, Y.Qian, R.R.Hudson et al., 2005 Interrogating multiple aspects of variation in a full resequencing data set to infer human population size changes. Proc. Natl. Acad. Sci. USA 102: Wakeley, J., 1999 Nonequilibrium migration in human history. Genetics 153: Wakeley, J., 2000 The effects of subdivision on the genetic divergence of populations and species. Evolution 54: Wakeley, J., 2001 The coalescent in an island model of population subdivision with variation among demes. Theor. Popul. Biol. 59: Wakeley, J., 2004 Metapopulations and coalescent theory, pp in Ecology, Genetics, and Evolution of Metapopulations, edited by I. Hanski and O. Gaggiotti. Elsevier, Oxford. Wakeley, J., and N. Aliacar, 2001 Gene genealogies in a metapopulation. Genetics 159: Wakeley, J., and J. Hey, 1997 Estimating ancestral population parameters. Genetics 145: Wakeley, J., and S. Lessard, 2003 Theory of the effects of population structure and sampling on patterns of linkage disequilibrium applied to genomic data from humans. Genetics 164: Watterson, G. A., 1975 On the number of segregating sites in genetical models without recombination. Theor. Popul. Biol. 7: Wegmann, D., M. Currat and L. Excoffier, 2006 Molecular diversity after a range expansion in heterogeneous environments. Genetics 174: Whitlock, M. C., and D. E. McCauley, 1999 Indirect measures of gene flow and migration: F ST doesn t equal 1/(4Nm 1 1). Heredity 82: Wilkinson-Herbots, H. M., 2008 The distribution of the coalescent time and the number of pairwise nucleotide differences in the isolation with migration model. Theor. Popul. Biol. 73: Wright, S., 1951 The genetical structure of populations. Ann. Eugen. 15: Wright, S. I., and B. S. Gaut, 2005 Molecular population genetics and the search for adaptive evolution in plants. Mol. Biol. Evol. 22: Communicating editor: N. Takahata

13 Supporting Information The Impact of Sampling Schemes on the Site Frequency Spectrum in Nonequilibrium Subdivided Populations Thomas Städler, Bernhard Haubold, Carlos Merino, Wolfgang Stephan and Peter Pfaffelhuber Copyright 2009 by the Genetics Society of America doi: /genetics

14 2 SI T. Städler et al. FIGURE S1. The same parameters were used as in Figure 1 of the main text. Here, we used an island model with 100 demes.

15 T. Städler et al. 3 SI FIGURE S2. The same parameters were used as in Figure 2 of the main text. Here, we used an island model with 100 demes.

16 4 SI T. Städler et al. FIGURE S3. The same parameters were used as in Figure3 of the main text. Here, we used an island model with 100 demes.

17 T. Städler et al. 5 SI FIGURE S4. The same parameters were used as in Figure 4 of the main text. Here, we used an island model with 100 demes.

18 6 SI T. Städler et al. FIGURE S5. The dependence of D T and D FL on the number of islands I in an island model. We fix τ /I = 0.02, β = 10 and 4N 0 m = 10. In addition θ I = 100 is held constantå

19 T. Städler et al. 7 SI FIGURE S6. The same parameters were used as in Figure 5 of the main text. Here, we used an island model with 100 demes.

20 8 SI T. Städler et al. FIGURE S7. Same as Figure 6 in the main text, but for D FL.

21 T. Städler et al. 9 S FIGURE S8. Same as Figure 6 in the main text, but for an island model with I = 400 islands.

22 10 SI T. Städler et al. FIGURE S9. Same as Figure S8, but for D FL.