Plant material and DNA isolation

Sequence data was generated from a sample of 32 haploid seed megagametophytes of loblolly pine (same plant material as in [14], see Table S2 from Supplementary Material). Note that these specimens are parents to a breeding population and are different than the set of samples used by Brown et al. [15], including several accessions from Florida not considered by [15]. For outgroup sequencing, two Scots pine (Pinus sylvestris) seed samples were provided by Dr. Päivi Komulainen and Dr. Outi Savolainen (University of Oulu, Finland). Scots pine was chosen as an outgroup since; i) its genetically well studied, with a genetic map and a mapping population, ii) has a comparative genetic map with loblolly pine, iii) Among the species that has comparative genetic maps with loblolly pine, it's the most distantly related species within the pine taxa. Total DNA was isolated using the QIAGEN Plant DNA extraction kit (Valencia, CA) in 96-well microtiter plate format, according to the manufacturer's protocol.

Candidate gene selection

Forty-one candidate genes were sequenced specifically for this study. For the references associating the indicated candidates to disease response from the literature see Supplementary Material- References S1. The selection of candidate gene loci was based on three major criteria:

(1) Positional candidates: Genes mapping near QTLs for cell wall chemistry [16]. These are the genes that had prior evidence from studies on other species to be involved in disease resistance response and located near QTLs for cell wall chemistry components such as lignin content, and cellulose content. (2) Expression candidates: Genes identified to have differential expression in northern blot or membrane array experiments [17]–[19] conducted to identify loci that are differentially expressed in loblolly pine seedlings infected by the two studied pathogens and uninfected controls or through electronic northerns between EST libraries generated from tissues under several different stress treatments and (3) Functional candidates: Genes that have been selected from the literature on model species whose homologs were identified in loblolly pine through homology searches within EST libraries available for loblolly pine and can be genetically mapped to a single locus in loblolly pine. Note that in this case the extent of the candidate gene list is limited to the coverage of the loblolly pine EST libraries. Initially, loci were selected from two main functional classes: genes annotated as metabolic enzymes and genes annotated as regulators of gene expression. However, upon data mining from public sequence databases, a small number of promoter-like intergenic sequences were identified as putative targets for some of the transcription factor families that were previously identified as candidate genes. These were also included in the target locus list (Table S3 from Supplementary Material).

PCR, sequencing and alignment

Conifer seeds' endosperm tissue, called megagametophyte is haploid. All the PCR and sequencing analyses was performed on DNA extracted from haploid megagametophyte tissue from the seeds. PCR and sequencing primers were designed and sequencing was performed as described in Gill et al. [20]. For PCR primer sequences see Table S4 from Supplementary Material. Sequencing reactions were run on ABI3730 sequencers at College of Agricultural Sciences, Genomics Facility (CGF) of UC-Davis. Consensus sequence from multiple reads of individual samples were generated from the trace files through base calling and multiple sequence alignment (MSA) by phred [21], phrap [22] and MACE-multiple alignment consed extensions (Gilliland & Langley, University of California, Davis), and visualized by consed [23]. All trace files were visually scanned for overlapping peaks as indicators of amplification from multiple loci (i.e. paralogs). Since the DNA used for amplification was haploid, any overlapping peaks in any sequence traces indicated amplification from multiple loci. Only contigs that consistently yielded sequences with higher than 40 phred quality score at all sites (according to the MACE alignment) were subjected to SNP calling. Furthermore, once a locus was identified to be putatively single locus, in cases when it was polymorphic between the parents of mapping populations, it was considered for genetic mapping in loblolly pine mapping populations [24] and checked for co-linearity and synteny between the loblolly pine and Scots pine comparative maps.

All loci were annotated by comparing MSAs with their respective template EST sequences, followed by manual editing according to the EST template, UniGene translations and known intron splicing sites. Sequences and alignments were submitted to NCBI-GenBank database with accession numbers EU392746-EU394132.

Exclusion of paralogous regions

In studies that heavily utilize PCR amplification for studying patterns of diversity and divergence, one serious weakness is the high incidence of amplification of high sequence similarity paralogous loci that may complicate the observations. As described in the previous section, we have tested the polymorphisms from each of the candidate genes, for their segregation between parents of two loblolly pine mapping populations, and one pair of parents from Scots pine mapping population, as well as a pair of mapping population parents for each of Monterey pine (Pinus radiata) and slash pine (Pinus elliottii) since each of these species have a comparative map generated via ESTP markers originally designed for loblolly pine. We were hoping to genetically map all these loci in all available mapping population to test for synteny and orthology across the loci amplified across species, since in case of a paralog amplification problem, segregation patterns are not expected to be robust. Unfortunately, not many of the loci included as candidates are segregating in mapping populations. Across all 71 loci screened for patterns of selection in this study coming from wood quality, drought tolerance and disease response projects, only 25 was successfully placed on the linkage map for loblolly pine. This low number is mostly due to SNPs not segregating between the parents, and assay design issues associated with designing genotyping assays for screening larger samples. As a result, 25 of the 71 loci mentioned in this study were genetically mapped to single genetic positions in mapping populations. We also have used FBRC-family based association population of loblolly pine (∼16 progeny per family on average) to check for deviations from Mendelian expectations for segregation of these and other loci. None of the loci displayed a significant deviation from expected Mendelian patterns in the FBRC-families for their segregation. (See Supplementary Material- Results S1 for data and details). Therefore, we concluded that it is unlikely that the patterns observed are due to amplification from paralogous copies of the genes either within or between species.

Fitting neutral demographic models

The site frequency spectrum (SFS) at each locus is the vector (x₁, x₂, … , x_n-1), where x_i is the number of polymorphisms observed in a sample of size n with i derived alleles and n-i ancestral alleles. However, for evaluating demographic models we did not distinguish the ancestral and derived alleles, even for loci for which an outgroup sequence was available. Thus, unless otherwise stated, all SFS analyses used the folded SFS, where the spectrum has n/2 frequency classes and i≤n-i indicates the minor allele frequency.

To estimate parameters for demographic models described below, we used coalescent simulations to draw random ancestral recombination graphs (ARGs) conditional on specific demographic parameter values and compute the likelihood of observing the experimental data given these ARGs using custom programs. Then, we optimized these parameters by searching for a configuration which produced the highest likelihood. To compute the likelihood given a particular ancestral recombination graph (ARG) and the scaled mutation rate Θ, the probability of observing an SFS is a product of Poisson densities(1)where the Poisson rate b_i•Θ is the sum of branch lengths in the ARG leading to a mutation of frequency i, times the scaled mutation rate Θ. As mentioned above, this likelihood formula depends on the demographic model and its parameters indirectly through the simulated ARG, so that all ARGs must be integrated over the space simulated to obtain the likelihood of the SFS given a particular demographic model. This integration is approximated by generating a large number of random ARGs conditional on the demographic model and averaging the likelihoods computed over this sample of ARGs. The likelihood can be computed to arbitrary precision by increasing the number of ARGs sampled. If M ARGs are sampled, the approximate likelihood formula for an individual locus is: (2)

We assume all loci analyzed here are effectively unlinked with no interactions between loci (a reasonable assumption given the low levels of LD among candidate genes in this dataset), so the likelihood over all loci is simply the product of individual locus likelihoods.

The likelihood formula above is critically dependent on the scaled mutation rate Θ that is estimated from the data. Watterson's moment estimator Θ_W [25] is biased under non-equilibrium demography, but it is easily shown and verified with simulations that under simple demographic scenarios such as those considered here, this estimate is biased by a constant factor which does not depend itself on Θ. This constant factor can be estimated simply by simulating under the demographic model of interest with a known Θ and comparing this value with Θ_W estimated from the simulated data. For our evaluations here, we used this procedure to estimate the constant factor and correct Θ_W estimated from the data for each specific model considered. Since we expect considerable variability at each locus with respect to this parameter, we computed an average estimate across coding and putative promoter loci. This distinction is necessary since promoter loci harbor substantially more diversity than coding loci (see Table S5 from Supplementary Material).

Not shown in formulae [1] and [2] is the dependence of the ARGs on the scaled recombination rate. For generating ARGs with the coalescent, we used the average scaled recombination rate (ρ = 0.023/bp) [26] estimated from the data using LDHat [27] as a point estimate for this parameter. The likelihood calculation depends only weakly on the scaled recombination rate, so long as it is not zero or near zero, so results are expected to be robust to any reasonable uncertainty in this parameter.

Previous work noted that glacial advances and retreats during recent ice ages restricted the habitable range of loblolly pine, causing fluctuations in census population size and most likely the effective population size [14], [15], [28], [29]. In order to generate a neutral expectation for the site frequency spectrum (SFS) at a locus, we must first determine if these suspected effective population size changes occurred and identify the time and magnitude of the change(s). We considered a simple model describing recent fluctuations in effective population size in two or three discrete epochs as well as the single epoch constant size (null) model (see Figure 1). Since previous studies have found little to no population structure across the extant range of this species [14], we did not investigate models incorporating structure or migration.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Figure 1. Bottleneck models tested.

N_{0 ,} N_B and N_A refer to current, during bottleneck and before bottleneck effective population sizes, while T_A and T_B refer to time in 2N₀ generations to the change from N_B to N₀ and from N_A to N_B , respectively.

https://doi.org/10.1371/journal.pone.0014234.g001

For the two-epoch and three-epoch demographic models, we searched the parameter space by evaluating the likelihood on a grid of points. To efficiently search this space and provide a finer density grid in regions with highest likelihoods, we first performed a search on a coarse grid of points over the widest sensible range of parameter values with M = 10,000 samples, then selected a much smaller range of parameter values based on the regions with highest likelihood on the coarse grid. This smaller range is then searched again with finer spacing between grid points, using M = 20,000 samples. Finally, since the highest likelihood models can differ only slightly in likelihood values and the evaluating equation [2] can result in slightly different values over multiple runs due to its stochastic nature, the highest likelihood models was evaluated again using M = 1,000,000 samples and the average likelihood over 10 such evaluations was taken. Likewise, to increase stability of the grid searches, likelihoods were evaluated three times in both the coarse and fine grid searches, and the average of the three evaluations used to determine the range of the finer grid and the selection of models for the final deep sample evaluation.

The ranges searched for the coarse grid were T_b ∈ [0.01, 4.0], T_a ∈ [0.01, 4.0], N_b/N₀ ∈ [0.01, 100.0], N_a/N₀ ∈ [0.01, 100.0] (see Figure 1 for description of model parameters). For all parameters on the coarse grid, 15 points at regular intervals spanning these ranges on log-scale were chosen, resulting in 225 two-epoch models and 30,875 three-epoch models. In both the coarse and fine grid searches, three-epoch models specifying T_a≤T_b were considered invalid and were not evaluated. The ranges for the fine grid search were T_b ∈ [0.01, 0.08], T_a ∈ [0.05, 0.5], N_b/N₀ ∈ [0.02, 0.4], N_a/N₀ ∈ [1.0, 10.0] for the three-epoch models and T_a ∈ [0.7, 1.8], N_a/N₀ ∈ [3.6, 30.0] for the two-epoch models. The finer search was conducted with 20 grid points at regular intervals on log-scale for each of the parameters, resulting in 273 two-epoch and 154,800 three-epoch models. Finally, based on the average likelihood of models evaluated in the fine grid search, all models within 4 log likelihood units of the maximum likelihood model were re-evaluated with a deeper sampling of M = 1,000,000 ARGs, and the average of 10 evaluations taken as the final likelihood value. The top competing models within 2 log likelihood units of the maximum likelihood model in each class are presented as the final results, indicating both the maximum likelihood model and a range of “nearby” models which collectively describe the uncertainty in parameter estimates.

To test significance of the three different classes of demographic models, we note that the constant size model is a specific case of the two- and three-epoch models, and the two-epoch model is a specific case of the three-epoch model. Thus, with nested models, we can formally test the significance of higher likelihoods in the more complicated models with a standard likelihood ratio test. For the comparison of the two-epoch with the constant size model, and the three-epoch with the two-epoch model, twice the difference in log likelihood is expected to be χ²-distributed with two degrees of freedom, under the null hypothesis that the model with more parameters does not significantly describe the data better than the simpler model.

Site Frequency Spectrum (SFS) diversity analysis

To assess whether individual loci were inconsistent with expectations under the best-fit neutral demographic model, coalescent samples of ARGs conditional on the best-fit model were generated placing mutations on the ARG conditional on the scaled mutation rate estimated for all loci within the coding and promoter locus classes, respectively. Then, the simulated site frequency spectra, corresponding summary statistics, and the likelihood of the simulated data were calculated. Empirical distributions of summary statistics, such as Tajima's D [30], Fu and Li's D [31] and the number of segregating sites, were compiled based on M coalescent simulations. Using these empirical distributions, the quantile (proportion) of the simulated values less than observed values can be calculated. Here, M = 100,000 samples under the best-fit demographic model were used to generate the empirical distributions.

McDonald-Kreitman Poisson Random Field (MKPRF) analysis

Information about average selective forces at a locus, over a longer period of evolutionary history, can be obtained from the level and patterns of nucleotide substitution between two closely related species. Under neutral evolution, polymorphisms represent a transient phase of the neutral substitution process, thus we should expect the ratio of substitutions between two species to polymorphisms within a species to be a function of the species divergence. Furthermore, if two distinct types of sites are both neutrally evolving, this ratio should be the same for each type. The standard McDonald-Krietman [32] test defines two types of sites within the same protein coding locus as silent or replacement based on whether or not the substitution or polymorphism alters the amino acid sequence. Under the assumption that silent sites are selectively neutral, a significant difference between the ratio of substitutions to polymorphisms for the two types of sites may indicate a locus is experiencing directional selection (too many replacement substitutions), or a strong selective constraint (too few replacement substitutions). This framework was generalized in the McDonald-Krietman Poisson Random Field (MKPRF) approach where information across multiple loci can be used to estimate global, locus-independent parameters such as the time of species divergence, allowing for more precise estimates of locus specific effects [33].

The MKPRF approach is based on a hierarchical Bayesian model which can describe not only the strength and direction of selection at individual loci but also an overall trend of selection across a group of loci. Briefly, the selection coefficient is defined for each replacement site within a locus, but modeled as independent, normally-distributed random variables with a common mean and variance. The program implementing the MKPRF approach provides a sample of the posterior distribution of the mean and variance of these random variables which is then summarized to describe the overall strength and direction of selection at each locus. In addition, these locus specific means and variances of selection coefficients are themselves treated as independent, normally-distributed random variables. These “hyper-parameters” describe the overall trend of selection across a group of loci indicating whether these loci as a group are experiencing positive or negative selection.

To perform the MKPRF analysis we used software provided by C. Bustamante, described in Bustamante et al. [33]. We distinguished the disease resistance loci sequenced specifically for this study from wood quality and drought tolerance candidate genes from previous studies [14], [15]. In addition, on the basis of preliminary results at the single locus level (see Figure 2), we subdivided the disease candidate loci into a conserved and non-conserved group (pr4.3, pr4.1, ldox-c and set-like-c; see Results and Supplementary Material Table S6). MKPRF was run with default parameters which specify diffuse, “uninformative” priors for the parameters of interest. MKPRF obtains samples from the posterior distribution of the parameters using the Monte Carlo Markov Chain (MCMC) approach. Ten independent chains were run for 10,000 “burn-in” iterations and then sampled every 10 iterations for a total of 10,000 samples for each chain. MKPRF outputs the Gelman-Rubin statistic for each parameter which in all cases was close to 1.0, indicating the chains have converged to the target posterior distribution. We summarized the posterior distributions of the selection coefficients for individual loci by the mean and 95% credible interval. The empirical densities obtained from 10,000 draws of the posterior distribution of the mean selection coefficient for groups of loci (see above) were directly plotted in Figure 2A. The input data can be found in Supplementary Material Table S6.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Figure 2. MKPRF analysis showing the posterior distribution of the mean selection coefficient across groups of loci (panel A), and the mean and 95% credible interval of the posterior distributions for per-locus selection coefficients (panel B).

A selection coefficient less than zero implies selective constraint at amino acid replacement sites while a coefficient greater than zero implies accelerated fixation of amino acid replacements.

https://doi.org/10.1371/journal.pone.0014234.g002

Species Isolation Model (IM)

We used the method of Wakeley & Hey [34] as implemented in WH software to estimate the parameters of a general species isolation model with parameters θ₁ = 4N₁µ of species 1, θ₂ = 4N₂µ of species 2, θ_A = 4N_Aµ of the ancestral population, and τ = 2µt, the time of the divergence between the species in units of N₁ generations. The IM model is based on counts of polymorphism specific to species 1 (S_x1), polymorphism specific to species 2 (S_x2), polymorphism fixed between the two species (S_F) and polymorphism segregating in both species (S_S). Once a general IM was fitted, we considered polymorphism types at each individual locus to assess whether observed patterns of polymorphism and divergence were consistent with expectations under this model. Similarly to the SFS analyses, coalescent simulations were performed to obtain the expected values under the fitted isolation model. Significance of deviations from these expectations for individual loci was assessed using the standard G-test (log-likelihood ratio test). Input data can be seen in Supplementary Material Table S7.

Loblolly pine demography and SFS analysis

We explored the parameter space for the two- and three-epoch models in a maximum likelihood framework and formally compared each epoch class with a likelihood ratio test (see Methods). Parameters for the highest likelihood models in each class are given in Table 1. The likelihood ratio test between the maximum likelihood three-epoch and two-epoch models was highly significant (p = 4.52×10⁻¹¹), showing that at least three periods of different population sizes must be considered to describe the data. This model describes a recent bottleneck in this species with about 85% reduction in effective population size, preceded by an ancestral epoch with an effective population size roughly four times the present epoch size. This is consistent with a population decline, possibly during a series of glacial advances during ice-age(s), followed by an incomplete recovery, possibly due to competition with emerging angiosperm species and a loss of range. Due to computational constraints and the exponential increase in the size of the parameter space with additional epochs, we did not consider more complex models of population size fluctuation. While it is likely that the true history of the species is more complex than the three-epoch model, this model may still be adequate to derive neutral expectations for present day population samples.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Table 1. Fitted model likelihoods for neutral, two-epoch and three-epoch demography models (see parameter meaning in Figure 1).

https://doi.org/10.1371/journal.pone.0014234.t001

Given the parameters for the three-epoch model fit to the data for all loci, we then considered whether any of the loci showed patterns of nucleotide polymorphism that were inconsistent with this neutral bottleneck expectation. The results for all loci are shown in Supplementary Material Table S8. At the α = 0.05 significance level, four loci (c4h1b, cad, erf1-like and ccoaoemt) showed an excess of mid-frequency alleles (higher than expected Tajima's D or Fu and Li's D), five (cyp450-like, c3h-f4r6, erd3, pr4.1 and myb3) showed an excess of rare alleles (lower than expected Tajima's D or Fu and Li's D), three had an excess of polymorphisms (agp4, wrky-like-2 and pr4.3) and six loci had fewer polymorphisms than expected (axr, gatabp1, caf1, erd3, paeoemt and pp2c). One locus (erd3) had simultaneously low diversity and an excess of rare alleles. When the significance threshold was raised to α = 0.01 only axr, c3h-f4r6, erd3, myb3 and pp2c were significant. Finally, only pp2c (less polymorphism than expected) and myb3 (excess of mid-frequency alleles) were significant when a (conservative) Bonferroni correction for multiple testing was applied (see Supplementary Material Table S8).

These observations may be consistent with balancing or diversifying selection in the case of excess mid-frequency alleles, and selective constraint or a recovery of diversity after a recent selective sweep for loci with an excess of rare alleles (see Supplementary Material Table S1). Low and high diversity loci may simply indicate a lower or higher mutation rate at these loci than the average mutation rate considered over all loci (see Methods). However, low diversity is also expected if there has been little time for recovery of neutral polymorphism after a recent selective sweep.

McDonald-Krietman Poisson Random Field analysis (MKPRF)

In our study, the posterior means of the selection coefficient distribution for each individual locus are of primary interest in order to evaluate the different selection hypotheses tested. Additionally, we asked whether disease candidate loci show a trend for selection that is distinct from other groups of candidate loci previously published [14], [15], since a central assumption for this work is that disease response loci are under stronger selection than most other loci in the genome.

MKPRF analysis was based on 32 of the disease response candidate loci in this study, 12 wood quality candidate loci [15] and 12 drought tolerance candidate loci [14], using publicly available (GenBank) spruce ESTs as an outgroup (Picea glauca and Picea sitchensis). It was immediately observed that the disease dataset contained four loci (pr4.3, pr4.1, ldox-c and set-like-c) that were strongly conserved (i.e. they had a highly significant deficit of amino acid replacements) and were poorly modeled as a single group together with the other disease candidates. We separated these loci into their own class and obtained the posterior distribution summaries shown in Figure 2A. Apart from the conserved pr4.3, pr4.1, ldox-c and set-like-c, results showed several other disease candidate loci with a significantly positive selection coefficient at α = 0.05. However, only gatabp1 was significant at α = 0.01 (Figure 2B). None of the non-disease candidate loci showed a significant positive or negative selection coefficient distribution.

When the different groups of candidate loci were compared, the distribution of mean selection coefficients was clearly different. The wood quality and drought tolerance candidate loci were not in general under positive or negative selection, but the non-conserved disease candidate loci showed a clear trend of positive selection overall. The conserved disease candidate loci (pr4.3, pr4.1, ldox-c and set-like-c) showed strong negative selection and were clearly a distinct group. Together, these results show that the disease candidate loci studied here as a whole are evolving in a distinct manner from either the wood quality or drought tolerance candidate loci, with some disease candidate loci individually showing significant evidence for adaptive evolution.

Species Isolation Model (IM)

For 24 of the disease candidate loci in this study and three of the drought tolerance candidates from previous work, we were able to amplify in single copy one or two gametes from the related pine species Pinus sylvestris (Scots pine). We noted that some loci harbored polymorphisms that were present in both species. This observation is unexpected since loblolly and Scots pine lineages diverged about 25 MYA [28], loblolly pine is found only in the southeastern United States while Scots pine's natural range is northern Europe and Asia, and there have been no reports of secondary contact between the species. One possibility is that close paralogs are being amplified in some individuals of each species. If this is the case, these polymorphisms identified would be paramorphisms rather than true SNPs. Paramorphisms do not necessarily exhibit Mendelian segregation; therefore, they won't be easily converted to informative genetic markers [29]. Acting on this expectation, we tested for Mendelian segregation of these loci within the two available mapping populations for loblolly pine with 90 progeny each, as well as families of the FBRC-family based association population of loblolly pine (∼16 progeny per family on average). Of the three genes that were identified to contain shared polymorphisms (lp5, pr4.1, wrky-like-1) lp-5 was successfully mapped to linkage group 3 using mapping populations and did not show a significant deviation from expected Mendelian patterns in the FBRC-families it was segregating in (lowest P = 0.087,χ² = 2.92(df = 1)). pr4.1 was not segregating in the mapping populations, however, when checked for their Mendelian segregation patterns between the parents and progeny of the FBRC-association population families, it also did not show a significant deviation from expected Mendelian patterns (lowest P = 0.10, χ² = 2.2 (df = 1)). wrky-like-1 displayed usually low values for the segregation ratios, displaying a marked decrease in heterozygote individuals detected. Regardless, all the p-values for each of 5 families that were segregating displayed P-values over the 0.05 threshold (lowest P = 0.072, χ² = 3.25 (df = 1)) (See Supplementary Material- Results S1).

Assuming, these polymorphisms are in fact true SNPs and not paramorphisms, it is of particular interest whether or not these shared polymorphisms have been preserved in both taxa as a result of the long-term action of balancing selection. With no reason to suspect gene flow between the two taxa, the expectation of shared polymorphisms can be estimated from the simple isolation model (IM) of speciation, as implemented in the WH program. Provided the isolation model of speciation is a good fit to the data overall, deviations from model expectations at individual loci would suggest population genetic forces other than mutation and drift are required to explain the observations.

First, the WH program was run to estimate overall population sizes and time of divergence. The WH test statistic computed by the program indicated a good overall fit of the isolation model with 61% of simulated datasets generating a higher than expected value of the test statistic. The time of divergence was estimated at 0.86N₁ generations with an ancestral population size 6.72 times the current size of loblolly pine. The Scots pine population size was estimated at 1.03 times the current size of loblolly pine. Second, simulated datasets were generated using these parameters to estimate an expected number of species specific polymorphisms, shared polymorphisms, and number of fixed differences for each locus.

Using a standard goodness-of-fit G-test, six loci (axr, gatabp1, lp5, pcna, sams2 and set-like-c) showed a poor fit to the isolation model at the 0.05 level (Bonferroni corrected) and five at the 0.01 level (Table 2). axr and gatabp1 have too few polymorphisms in loblolly pine and too many fixed differences. This is in agreement with the low nucleotide diversity found in the SFS analysis for these two loci. The excess of fixed differences indicates a rapid rate of substitution and is consistent with directional selection. pcna has too many polymorphisms in loblolly pine, and sams2 and set-like-c have too many polymorphisms in Scots pine; all three loci also appear to have too few fixed differences between the species. Too few fixed differences could be explained by balancing selection acting in only one of the species, reducing the rate of fixation between species while increasing or maintaining polymorphism within that species. Lack of balancing selection in the other species would allow the ancestral alleles to be lost by drift and shared polymorphisms would not be maintained. Alternatively, purifying selection against new variants could explain the reduced level of fixed differences, but this is not consistent with elevated polymorphism levels in either species or the fact that the SFS analysis did not reveal any excess of low frequency polymorphisms for these three loci. Conversely, the MKPRF analysis found that set-like-c was highly conserved, which is consistent with purifying selection.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Table 2. Isolation model (IM) for Scots and loblolly pines.

https://doi.org/10.1371/journal.pone.0014234.t002

Of the three loci with shared polymorphisms, only lp5 showed a significantly poor fit to the isolation model, with an excess of shared polymorphisms, excess of polymorphisms in loblolly pine, and a lack of fixed differences. This seems consistent with balancing selection maintaining ancestral alleles in both species as well as increasing diversity within loblolly pine. However, an excess of mid-frequency alleles, which might be expected under some balancing selection scenarios, was not observed in the SFS analysis.

The neighbor-joining tree plots of the polymorphism and divergence data from the loci are available as Supplementary Material in file Results S2.