We used a random subset of 1,000 markers to conduct a Bayesian inference of population structure, which revealed a strong peak in ΔK for K = 3 (ΔK for k2 = 84.6, ΔK for k3 = 713.1 and ΔK for k4 = 12.25) largely corresponding to the species H. annuus, H. petiolaris and H. debilis/praecox (Figure 2). We analyzed population structure using three different sets of 1,000 random markers and obtained the same strong clustering results each time. One individual (btm15–2) clearly stood out as an early generation H. debilis-H. annuus hybrid in all runs conducted (Figure 2). We removed this individual from further analyses of genetic, protein coding and gene expression divergence.

Given that no distinct genetic clusters between H. debilis and H. praecox were identified at K = 3 (Figure 2) and the analysis of population structure with k = 4 did not reveal a clear split, all H.praecox individuals were subsumed within H. debilis for further analyses of genetic, protein coding and gene expression divergence. We also calculated the average mean F_STper gene as 0.25, 0.23 and 0.28 for H. annuus-H. debilis, H. annuus-H. petiolaris, H. debilis-H. petiolaris comparisons respectively (Table 1).

We obtained open reading frames (regions greater than 300 bp without stop codons) for 12,310 of the 16,312 reference sequences. We then used paml to calculate d_N and d_Sin a maximum likelihood framework. Mean d_N/d_Sratios were 0.203, 0.203 and 0.205 for the H. annuus-H. debilis, H. annuus-H. petiolaris and H. debilis-H. petiolaris comparisons, respectively (Table 3). Lastly, d_N and d_Swere strongly correlated (r2 values between d_Nand d_Swere 0.36, 0.33, 0.27 for H. annuus-H. debilis, H. annuus-H. petiolaris and H. debilis-H. petiolaris comparisons, respectively).

We then tested whether the ratio of non-synonymous to synonymous fixed differences was greater than the non-synonymous to synonymous polymorphism ratio using a g-test. For all comparisons, there was a significant excess of non-synonymous fixed differences (g-test < 0.0001 in all cases, Table 3). The genome wide proportion of base substitutions fixed by natural selection (alpha) was observed to be 0.28 for H. annuus-H. debilis, 0.23 for H. annuus-H. petiolaris and 0.25 for H. debilis-H. petiolaris. When focusing on the most abundant GO categories (comprising at least 50 genes), we found evidence that the proportion of adaptive evolution (alpha) varied significantly with the class of biological processes considered. In Table 4, we present the five GO categories with the highest alpha value in each comparison. These GO categories comprised a number of biological processes. In only one case, however, was the same GO category (response to biotic stimulus) present in the “top five” in all three species pairs. Other GO categories with very high alpha values in at least one comparison included flower development, glycolysis, defense response to fungus, lipid catabolic process, and lipid metabolic process.

We conducted partial correlation analyses among different factors which have been previously shown to correlate with variation in rates of protein evolution. First, we only analyzed d_N/d_S ratios for open reading frames in which both d_N and d_S were different from zero (4,908, 4,161 and 4,196 ORFs for H. annuus-H. debilis, H. annuus-H. petiolaris and H. debilis-H. petiolaris comparisons, respectively). Then, we noticed that for all three interspecific comparisons, correlation coefficients were similar. Therefore, in order to simplify the interpretations, we have presented the data as a mean of the three comparisons (Figure 3, but see also Figure 4, Figure 5, Figure 6 for correlations per species pair). Rates of protein divergence (d_N/d_S) were negatively correlated with gene expression level (partial cor. coef. = −0.03) and positively with gene expression specificity (0.06), consistent with results from a broad range of taxa [6,8,28] and positively correlated with overall sequence divergence (F_ST, partial cor. coef. = 0.04). Among other notable results predicted from previous studies, we found a strong and significant negative correlation between expression level and specificity (−0.15). Nucleotide diversity (pi) was strongly negatively correlated with sequence divergence (−0.27) and surprisingly, positively with expression level (partial cor. coef. = 0.48). We also found that sequence divergence (F_ST) did not correlate with expression divergence.

Our approach using a reference transcriptome of 16,312 contigs allowed to survey a substantial fraction of all genes expressed and capture genome wide signals that would likely go undetected when using a handful of anonymous markers or a candidate gene strategy. Nearly every gene in our reference set had a least two sequences align to it and could therefore be considered expressed (15,756 out of 16,312 contigs) in the sequenced samples. The expression cut-offs commonly employed for microarrays or RT-PCR has created a perception that genes with a signal intensity below a certain ambient noise are not expressed. However, this perception probably relates more to the resolving power of the method rather than a real property of the gene; genes usually considered as silent are probably often expressed at a very low level making expression differences more subtle than originally perceived [29]. Conversely, we cannot rule out this possibility in some cases of uncertainty in sequence or alignment; for this reason genes with less than two aligned reads were removed from the data set. The fact that only about 50% of cleaned sequences aligned to our reference indicates that a significant fraction of the transcriptome remains un-assembled and/or that some reads are too divergent from the reference or represent chimeric sequences. Alternatively spliced genes and genes filtered from the reference dataset because they were inferred to represent families of close paralogs might also explain some of this missing fraction. In RNAseq studies, another source of bias can come from aligning reads from other species to a reference composed of a single species, here H. annuus. Fortunately, this does not seem to be a problem given the alignment parameters we chose and the fact that the same percentages (52%) of cleaned reads (Table 2) were aligned in all three species.

Given that the cultivated sunflower genome is nearing completion [30], there will soon be a more accurate reference to align sequences against. A fully sequenced genome will permit more accurate identification of gene space, alternative splicing events, untranslated 5' and 3' regions and intron/exon boundaries. In addition, we are concurrently developing reference transcriptomes for several closely related sunflower species including H. petiolaris and H. debilis. This will allow us to investigate orphan genes that are present in either H. petiolaris or H. debilis but have no detectable homolog in H. annuus. Such genes are rare, but may play an important role in generating species-specific evolutionary novelties [31]. Clearly, we have only begun to understand the evolutionary complexity underlying both sequence and expression divergence.

Annual species of sunflowers in the genus Helianthus vary greatly in their geographic distribution and evolutionary origin. Based on Bayesian clustering of nuclear markers spanning the whole genome, we found a clear distinction between H. debilis, H petiolaris and H. annuus individuals. However, structure often does not distinguish hierarchical sub-structuring and consequently, may only reveal the top level of population differentiation. Our study was not intended to resolve the phylogeny of Helianthus, as this would have needed much more extensive sampling. However, our results corroborate previous phylogenetic work [32,33] that places H. praecox within H. debilis.

As a whole, genetic divergence was surprisingly similar among all three species comparisons (Table 1) despite the fact that the divergence of H. annuus with H. petiolaris and H. debilis clearly precedes the H.petiolaris-H. debilis divergence [33]. Hybridization between recognized species is known to participate in shaping the evolutionary diversity of North American sunflowers [20,34,35] and probably explains the patterns observed here. H. petiolaris and H. debilis are allopatric sister species, but both hybridize with the sympatric but evolutionary more distant H. annuus. H. annuus and H. debilis are known to have a long history of introgression in Texas, where these individuals were collected [20,35]. Similarly, H. petiolaris forms numerous hybrid zones with H. annuus throughout their ranges [34]. As such, these more distantly related species have exchanged genes across much of the genome while remaining morphologically and ecologically distinct [16], presumably accounting for lower than expected average F_ST values.

The longest open-ended ORFs (minimum length of 300 nucleotides) were kept as the most probable translated region of the gene. SNPs within these ORFs were considered as coding sites as thus it could be assessed whether these were synonymous or non-synonymous. This approach (as compared with a blastx approach to known protein sequences from model systems) has the advantage of identifying genes that have no detectable homolog present in public databases. Such orphan genes are potentially rare, but may play an important role in generating species-specific evolutionary novelties [31].

Our results are consistent with several expectations based on previous research [14]. First, mean d_N/d_S ratios were close to 0.2 for all three comparisons and few genes were identified with values greater than one, signaling strong constraint on amino acid replacements. This is consistent with the results of Yang and Gaut [28] who analyzed protein coding evolution between Arabidopsis thaliana and A. lyrata and estimated d_N/d_S at 0.203. Many studies have documented a positive correlation between d_N and d_S across genes, similar to that identified here [14]. The reason for this correlation is not well understood, although it may be caused by processes such as strong purifying selection towards translational efficiency or variation in mutation rates and recombination rates along chromosomes, which would affect both d_Nand d_S in a similar fashion.

Based on calculations of the genome wide proportion of base substitutions fixed by natural selection (alpha), we conclude that a significant proportion of the most divergent SNPs represent selectively advantageous mutations: in all cases the ratio of non-synonymous to synonymous fixed differences was significantly greater than that for non-synonymous to synonymous polymorphisms (alpha, Table 3). By focusing on specific gene categories and examining patterns of polymorphism and divergence, we were able to provide additional evidence of adaptive evolution (Table 4). Gene categories exhibiting high alpha values included a number of biological processes that have previously been reported as undergoing positive selection in sunflowers, including defense response to fungus, flower development, lipid catabolic process, lipid metabolic process, and response to biotic stimulus [36,37,38,39,40]. Interestingly, we found one category, response to biotic stimulus, showing parallel levels of elevated alpha among species pairs. This is important given that parallel trends add stronger evidence for the role of selection in driving divergence for these particular traits [41]. Thus, natural selection imposed by other organisms appears to play an important role in driving protein evolution in wild sunflowers. However, we also acknowledge the dangers of story telling based on GO analysis and these trends should be confirmed before any strong conclusions can be reached [42].

To understand the underlying causes of evolutionary rate heterogeneity among genes, researchers have investigated correlations between evolutionary rates and various genomic parameters [1,8,28]. In accordance with previous studies (e.g., [28]), we found a significant negative correlation between d_N/d_S ratios and expression level, with genes that are highly expressed being more constrained in terms of protein evolution. Likewise, d_N/d_S ratios were positively correlated with expression specificity. Again the leading explanation for a positive correlation is that widely expressed genes are more constrained in protein sequence evolution because they are probably involved in multiple biochemical pathways and face multiple selective environments in the different tissues they are expressed in [5]. Finally, we found a strong and significant negative correlation between expression level and specificity. Highly expressed genes are usually involved in a larger number of biochemical processes and expressed in a greater number of different tissues than lowly expressed genes. Since specificity was determined based on relatively shallow (44,000 reads) sequencing, lowly expressed genes may appear to have high specificity due to the stochastic lack of sampling in some libraries. At the same time, it is likely that with enough power (i.e., enough sequencing), we may find that all genes are expressed nearly everywhere (i.e., expression is leaky, see [29]). Therefore specificity may eventually need to be redefined or at least measured in independent samples, as in the present study.

All the trends reported above are consistent with theory and results from a broad range of taxa [1,2,3,5,7,28,43]. Interestingly however, the correlation coefficients reported here are generally lower than those observed in previous studies. This is probably due to the more recent divergence of sunflowers compared with the more distantly related taxa studied previously. For instance, the sunflowers species analyzed here have diverged less than two million years ago [19], compared with about eight MYA in the Arabidopsis thaliana-lyrata comparisons [28], six MYA for the Drosophila melanogaster group (but with much faster generation times than sunflowers [1,44], and about 75 MYA in the case of the mouse and human lineages [5]. As recently pointed out in a study quantifying protein coding evolution broadly across the Asteraceae family [13], it appears that, at least to some degree, macroevolution looks very much like “repeated rounds of microevolution” [45]. Given that our results are also qualitatively similar to previous studies conducted at a more ancient timescale, we argue that macroevolutionary trends are governed by the same principles of microevolution. Yet, the importance of the genomic factors discussed above is possibly reduced relative to the idiosyncratic effects of drift and local variation in selection pressures when analyzed over shorter timescales [13].

Unlike previous studies, we also observed several new and sometimes unexpected associations. For one, we found that d_N/d_S ratios were positively correlated with overall genetic divergence (F_ST). Given that high values of d_N/d_S and F_ST can result from positive selection, this association could be expected. Among other noticeable results, we did not find a significant correlation between genetic divergence (measured as F_ST, or d_N/d_S ratios) and expression divergence. After observing no correlation in yeast between d_N and expression divergence, Tirosh and Barkai [11] suggested that certain genes are more sensitive to changes in coding sequences, whereas others are more sensitive to changes in non coding regions regulating expression. Consequently, an absence of correlation between patterns of expression and genetic divergence, as we identified in the present study, suggests a decoupling of evolutionary forces shaping expression and genetic divergence.

Unexpectedly, nucleotide diversity (pi) was strongly correlated with expression level in Helianthus. With RNAseq, the confidence with which nucleotide variants are identified is a function of coverage, which itself represents expression. As a consequence, greater coverage (expression) increases the probability of identifying both alleles at heterozygous loci and may inflate the total number of variants [46]. Analyzing only the subset of genes with d_N and d_S greater than 0 should ensure that most low coverage genes are removed (mean expression value for all genes in dataset = 365, compared with mean expression of genes used in partial correlation analysis = 665). Using an even more stringent criterion on the minimum expression of genes (for example, minimum expression threshold of 100 instead of now 14) does slightly reduce the correlation between pi and expression level but does not substantially affect the other partial correlations reported in Figure 3. Thus, our conclusions hold when using more stringent filtering procedures. Lastly, nucleotide diversity (pi) was also strongly negatively correlated with F_ST. This is not unexpected and likely pertains to the definition of F_ST itself ([heterozygosity between-heterozygosity within]/heterozygosity between).

Essentially all previous studies looking at the genomic determinants of sequence evolution have employed one dataset comprised of sequenced genes, a separate dataset for calculating gene expression (i.e., microarray data often measured in different individuals or even populations) and another dataset to identify genomic context (i.e., an annotated genome). In addition, previous studies have mainly estimated genetic divergence in terms of nucleotide substitutions between distant species, rather than F_ST, since the latter will be close to one in genetically distant species. Because of the independence between the expression and sequence data sets in these previous studies, the apparent ascertainment bias detected in the present study is avoided. Thus, while RNAseq increases the precision of expression analyses, and ensures that the same pairs of genes are compared for sequence and expression analyses, the approach is not without drawbacks.