We collected and analyzed sequence data from fragments of hundreds of expressed paralogs from multiple species with an aim of teasing apart early from later mutations in the protein coding region of persistent paralogs generated by WGD (Fig. 1). In Xenopus, a concatenated analysis of 80,856 base pairs (bp) of expressed paralogs indicates that synonymous substitutions per synonymous site (Ks) between X. laevis paralogs (XLα and XLβ in Fig. 1B) is 0.2111, and Ks between the alpha paralogs of X. laevis and X. borealis (XLα and XBα in Fig. 1B) is 0.1393. This suggests that Ks between paralogs in the "early" stage of duplicate gene evolution is up to 0.0718, depending on the location of node 2 in Fig. 1B. Most synonymous divergence between paralogs therefore accumulated after tetraploidization in Xenopus (see Additional file 1), which occurred roughly 20 to 40 million years ago 32 or maybe more 39. Silurana allotetraploids are about half as old 39.

After allopolyploidization, these paralogs were rapidly (immediately or soon after WGD) subjected to strong purifying selection. The level of purifying selection, while relaxed relative to singletons 3241, did not vary substantially between early and later stages of duplicate gene evolution.

More specifically, a combined analysis of thousands of codons from hundreds of expressed paralogs from X. laevis, a X. borealis ortholog, and a S. tropicalis ortholog, indicates that a more parameterized model of sequence evolution with a higher Ka/Ks ratio during the early stage of duplicate gene evolution than the later stage is not preferred (P = 1.00, Table 1, Fig. 1B). In fact, a branch-specific model of evolution indicates that the estimated Ka/Ks ratio in the early stage of duplicate gene evolution is slightly lower than in the later stage (Table 1). When these data were partitioned by gene fragment the results were the same – there also was not a significant difference in the Ka/Ks ratio at the early compared to the later stage of duplicate gene evolution (Table 1). Additionally, a model in which the Ka/Ks ratio of the early lineage is allowed to be lower than one is significantly better than a model in which this rate ratio is fixed at the neutral expectation of one (P < 0.00001, Table 2) and this analysis also produced the same result when the data were partitioned by gene fragment (Table 2). Similarly, branch-site models recover a higher proportion of positively selected sites in the later lineage (0.00893%) than the early lineage (0.00061%; data not shown).

Tests of the individual loci have low power because many are small fragments (see Additional file 2). Nonetheless, analyses of 660 fragments from 350 individual loci echo the results of the analyses of combined multi-locus data. The distribution of Ka/Ks ratios in the early and later stages of duplicate gene evolution is similar (Fig. 2A) and more fragments have a significantly higher Ka/Ks ratio at a later stage (8 fragments) than at an earlier stage (6 fragments), and this difference is not significant (χ² test, P = 0.997). Additionally, the number of fragments with a higher Ka/Ks ratio in the early stage than the later stage (significant or not) was lower (156 fragments) than the alternative (262 fragments; P = 1.0; see Additional file 2). That Ks in the early stage of duplicate gene evolution was similar to or lower than in the later stage (Fig. 1, see Additional file 3), indicates that sampling bias of synonymous substitutions 3243, if present, would bias our analysis of individual fragments towards detecting a higher Ka/Ks ratio in the early stage, which is not what we observed.

The neutral expectation (Ka/Ks equal to one) is significantly rejected in the early lineage of 62 out of 136 individual loci with more than 200 bp (see Additional file 2), and when this ratio is estimated for the early lineage, only 7% of them have an estimated Ka/Ks ratio above one. Taken together, these results indicate that purifying selection was as strong, if not stronger on these duplicates in the early stage of their evolution compared to the later stage.

Early neofunctionalization could potentially result in no difference between the Ka/Ks ratio in the early and later stages of duplicate gene evolution if genes in the early stage experience either positive selection or purifying selection, whereas genes in the later stage experience either relaxed purifying selection or purifying selection. While we can not rule this possibility out because positive selection and relaxed purifying selection both increase the Ka/Ks ratio, a regression of Ka/Ks to Ks for each fragment in the early and later stage of evolution indicates that (positive + relaxed purifying) selection is less prevalent in the early stage than the later stage (Fig. 2B).

New function may be achieved by "radical" substitutions – replacement of one amino acid with another that has very different physical properties 2444. While this is certainly not a requirement for new function to evolve, we nonetheless explored this possibility using a Bayesian approach to estimate the number and frequencies of elemental substitutions – the 75 amino acid substitutions that can occur via a single nucleotide change – at an early and a later stage of duplicate gene evolution, and also in the diploid lineage (see Additional file 3). Results indicate that elemental substitutions were not more radical in an early stage (Mantel Z statistic = 2.4119) than in a later stage (P = 1.0000). In fact, radical substitutions were slightly more prevalent in the later lineage (Mantel Z statistic = 2.4680). Elemental substitutions also were not significantly more radical in the entire X. laevis paralog α lineage (between node 1 and XLα of Fig. 1B, Mantel Z statistic = 2.43823) than in the diploid lineage (between ST and node 1 in Fig. 1B; Mantel Z statistic = 2.3920, P = 0.1396). Similar results are obtained when the radicalness of elemental substitutions is categorized according to alternative criteria [data not shown; 45].

Simulations were performed to test whether ancestral bias toward more conservative substitutions in the early stages of duplicate gene evolution could explain these results, but this was not the case. Simulated elemental substitutions from a reconstructed ancestral sequence were not more conservative in the early stage of duplicate gene evolution than the later stage (P = 0.6529). As expected, these simulations, which were not under purifying selection, were significantly more radical than the observed data (P < 0.001).

We performed additional analyses to address various concerns about the sequence dataset from X. laevis, X. borealis, and S. tropicalis. One consideration is that differences or changes in population size could affect the Ka/Ks ratio because slightly deleterious nonsynonymous substitutions are more likely to fix when the effective population size is small. Based on the geographic distribution and molecular diversity of mitochondrial DNA 39, the effective population size of X. borealis is smaller than that of X. laevis. However, we found that the Ka/Ks ratio of X. laevis paralogs during the later stage was slightly higher (0.1555) than the corresponding lineage of X. borealis (0.1338). This discrepancy was not significant in a two-sided test (P = 0.1790) or in a one-sided test because we expected the ratio to be larger in X. borealis (P = 1.0). To ensure that we were comparing ratios in expressed duplicates in both species, we included in this comparison only those data for which expression of both paralogs of both species was confirmed (37,194 bp). We note that more substitutions of both types occurred in X. borealis suggesting that the overall rate of evolution may be slightly higher in this species. A lack of significant difference in the Ka/Ks ratio suggests that the difference in effective population size between X. laevis and X. borealis had a negligible impact on the Ka/Ks ratios of many of their orthologs.

A second consideration stems from the possibility that a substantial portion of the early lineage of duplicate gene evolution evolved in a diploid (between nodes 1 and 2 in Fig. 1B) as a result of the putative allopolyploid origin of the ancestor of Xenopus tetraploids. Because the Ka/Ks ratio of clawed frog paralogs is slightly higher after genome duplication than before it 3241, the Ka/Ks ratio of this entire branch (between nodes 1 and 3 in Fig. 1B) could be lower than the Ka/Ks ratio of the portion of this branch that evolved after duplication (between nodes 2 and 3 in Fig. 1B). To explore this issue, we analyzed expressed sequences from another dataset derived from S. tropicalis and two closely related tetraploids (9717 bp). Similar to the analysis of X. laevis and X. borealis paralogs, the branch-specific tests of Silurana paralogs do not provide evidence for an increased Ka/Ks ratio in an early stage (between nodes 7 and 8 in Fig. 1C) versus a later stage of duplicate gene evolution (between node 8 and EPα in Fig. 1C; P = 1.0; Table 1), nor an increased frequency of radical amino acid substitutions at an early stage of duplicate gene evolution (Mantel Z statistic = 2.7193) compared to a later stage (Mantel Z statistic = 2.1991, P = 0.0882). Simulations indicate that the early stage of duplicate gene evolution in Silurana was not significantly biased towards more conservative substitutions (P = 0.5651), the branch-site test recovers no evidence in the concatenated data for positively selected sites in the early branch (although it does on the later branch; data not shown), and the partitioned branch model analysis recovers the same results as the concatenated branch model (Tables 1 and 2). Also similar to the analysis of X. laevis and X. borealis paralogs, the branch-specific tests of Silurana paralogs illustrate that functional constraints during the early stage of duplicate gene evolution were significantly below neutral expectations (Table 2).

A third consideration is that allotetraploidization of the common ancestor of Xenopus tetraploids occurred immediately before the first speciation of this ancestor (in other words that the time between nodes 2 and 3 in Fig. 1B is very small). If this were the case, then it would be more informative to compare an "intermediate" stage of duplicate gene evolution – a period after the first tetraploid speciation in Xenopus but before subsequent tetraploid speciations (i.e. between nodes 3 and 4 in Fig. 1D) – to a later stage of duplicate gene evolution – after an even more recent tetraploid speciation event (between node 4 and XLα in Fig. 1D). This issue was addressed with additional sequences (6966 bp) from the tetraploid species X. gilli and X. muelleri that made possible the further dissection and hypothesis testing of the temporal dynamics of evolution after duplication (Fig. 1D). Based on their close phylogenetic relationships 223946, we used X. gilli when we knew both X. laevis paralogs were expressed, and we used X. muelleri when we knew both X. borealis paralogs were expressed. Similar to the other analyses, this comparison revealed that the Ka/Ks ratio is not significantly higher in the intermediate stage compared to the later stage of duplicate gene evolution (P = 0.16; Table 1) and that the frequency of radical amino acid substitutions at the intermediate stage of duplicate gene evolution (Mantel Z statistic = 2.4073) is not significantly higher than at a later stage (Mantel Z statistic = 2.0645, P = 0.0887). Simulations again confirm that the intermediate stage was not significantly biased towards more conservative substitutions (P = 1.0000), the branch-site test recovers no evidence in the concatenated data for positively selected sites in the early branch or the later branch (data not shown), and the partitioned branch model analysis again recovers the same results as the concatenated branch model (Tables 1 and 2). These additional analyses thus provide strong support that purifying selection acted rapidly – within millions of years – and persistently – over tens of millions of years – after WGD in clawed frogs.

We used microarray data to compare expression profiles from five developmental stages and adult tissue types (treatments) of hundreds of paralogous pairs generated by WGD. Our analyses included developmental treatments from four distinct developmental stages (egg, tadpole stage 11, tadpole stage 18, and adult). Unlike the egg and tadpole stages, however, the adult stage is represented by data from each type of gonad instead of the entire individual. Because the primordial germ cells appear long after the tadpole stages that we assayed [stage 4447], these data provide a coarse perspective on spatial expression in four distinct tissue types: undifferentiated egg, pooled embryonic tissue (which do not have developed gonads), adult testis, and adult ovary.

We performed a power analysis to explore the possibility that cross-hybridization of non-target paralogs could affect the inference of paralogous expression profiles. We compared results from (a) a low paralog specificity analysis that included all probes on the microarray, including ones that cross hybridize to both paralogs, (b) a medium paralog specificity analysis that excluded probes whose sequences cross-hybridized to both paralogs, and (c) a high specificity analysis that excluded probes having up to three mismatches with a nontarget paralog. Additionally, we used two intensity thresholds, "standard" and "conservative", as a basis for the detection of expression of each paralog in each treatment (see Methods).

Qualitative comparisons across this developmental series and these tissue types indicate that the bulk of paralogous expression divergence after WGD in clawed frogs is on a quantitative rather than a temporal dimension (Figs. 3, 4). This would be expected if these paralogs were expressed in a highly specific manner in only one of the developmental stages or tissue types that we analyzed. However, many of these paralogs were expressed in multiple tissue types and multiple developmental stages. Consider for instance the 841 paralogous pairs for which the presence/absence expression profile of each paralog was identical in the medium and high paralog specificity analyses (Fig. 3). In the medium specificity analysis at the standard threshold, 94% of these paralogous pairs were both expressed in at least two treatments and 75% were both expressed in all five treatments.

When both paralogs are expressed, comparison of their expression profiles can indicate either that (a) both are expressed at the same developmental stages and tissue types (identical spatial and temporal expression), (b) the profile of one paralog is a subset of that of the other one (overlapping spatial and temporal expression), or (c) both paralogs have distinct components to their expression profiles (distinct spatial and temporal expression). In the microarray expression data from X. laevis, when expression of both paralogs was detected, almost all pairs had identical or overlapping expression profiles in terms of the developmental stages and tissue types in which expression was detected (Fig. 3). This was true regardless of how conservatively we scored presence/absence of expression or the specificity of the probes on the microarray. Only 2–7% of these pairs included paralogs that both had a unique expression profile wherein one paralog is expressed at a developmental stage or a tissue type where the other one is not expressed, and vice versa (Fig. 3).

In contrast to the overall similarity in the developmental timing and locations of paralogous expression, quantitative aspects of a high percentage of paralogous pairs have diverged substantially (Fig. 4). In the medium paralog specificity analysis for example, 62% of the paralogous pairs had a Pearson correlation coefficient that was below 0.866, a value below which were 95% of the correlation coefficients between non-paralogous genes. 27% of the paralogous pairs had a correlation coefficient below 0.5. Similar proportions were recovered in the high paralog specificity analysis (results not shown). At the end of this extreme, 0.3% of the paralogous pairs (3 pairs) in the medium paralog specificity analysis had a correlation coefficient that was more negative than -0.861, a level below which were only 5% of the correlation coefficients of the non-paralogous expression profiles. These three paralogous pairs are expressed in all treatments according to the standard detection threshold and have the following accession numbers (NM_001092603 and NM_001091285, NM_001091759 and NM_001093475, and NM_001091931 and NM_001094047). Their annotations are rudimentary, but the first pair may be involved with RNA splicing and the third pair has sequence similarity to collagen alpha(1) precursor. The normalized expression level of each pair indicates that in most of these treatments, the expression of one paralog is above the median expression level of that paralog across the five treatments whereas the expression of the other paralog is below it.

We compiled sequences of expressed paralogs of X. laevis from Genbank and various publications324142 and aligned them with orthologs from the S. tropicalis genome assembly 4.1. 454 pyrosequencing was used to obtain sequences of fragments of expressed paralogs of X. borealis from testis cDNA and contigs were assembled from these data using BLAST 80 and ALIGN0 81 from the FASTA 2.0 package 82, and manual alignment in MacClade 83. Manufacturer protocols were followed to isolate RNA using an RNA extraction kit (Qiagen), to prepare cDNA using BD SMART PCR cDNA synthesis kit (Clontech), and to normalize the cDNA using the Trimmer cDNA normalization kit (Evrogen JCS). Additional targeted sequencing of paralogs from X. laevis, X. borealis, X. gilli, X. muelleri, S. epitropicalis, and S. new tetraploid was performed by co-amplifying portions of these paralogs from cDNA from a variety of tissues (blood, heart, brain, testis, liver, muscle). Portions of individual paralogs were then cloned with the TA cloning kit (Invitrogen) and sequenced. These data are deposited in Genbank (see Additional file 2).

Using a combination of targeted amplification, cloning, and sequencing of cDNA, 454 pyrosequencing of cDNA, and database searches, 80,856 bp were collected from 660 fragments of 350 expressed paralogous pairs from the tetraploid X. laevis, one expressed paralog from the tetraploid X. borealis, and an ortholog from the diploid S. tropicalis. An additional 9,717 bp were sequenced from portions of thirteen expressed duplicated loci of the tetraploids S. epitropicalis and S. new tetraploid, and 6,966 bp were sequenced from portions of nine expressed duplicated loci of the tetraploids X. muelleri or X. gilli. To minimize analysis of paralogs whose evolutionary history may have been homogenized by gene conversion or recombination, we excluded from our analysis sequences with signs of these phenomena (see Additional file 1).

Because data were usually obtained from only one expressed X. borealis paralog but two X. laevis paralogs, most of our molecular analyses focused on molecular evolution of one X. laevis paralog – the "α" paralog (Fig. 1B). This is because, without evidence of expression of the other X. borealis paralog, we do not know whether X. borealis paralog α is still an expressed duplicate, and we also cannot determine at what point after duplication nonsyonymous substitutions occurred in the other X. laevis paralog – paralog "β" (Fig. 1B). Phylogenetic methods (maximum parsimony, maximum likelihood) were used to identify to which one of the expressed X. laevis paralogs that the X. borealis paralog was most closely related.

To test whether the rate ratio of nonsynonymous to synonymous substitutions per site (hereafter the Ka/Ks ratio) differs at early versus later stages of duplicate gene evolution, the likelihood of alternative models of branch-specific evolution (Fig. 1) was calculated using PAML version 3.15 84. This analysis was performed on concatenated datasets and with the data partitioned by gene fragment. We also used the branch-site test for positive selection [test 2 in 85] to test whether there were more sites under positive selection at an early stage compared to a later stage of duplicate gene evolution. In addition, we tested for significant departure from neutrality by comparing a model in which the Ka/Ks ratio of the early lineage was fixed at one and another Ka/Ks ratio was estimated for all other branches, to a model in which one Ka/Ks ratio was estimated for the early branch and another ratio was again estimated for all other branches. For each comparison, significance of the more parameterized model was evaluated with a χ² test. Note that, as a result of a suspected allotetraploid origin of the ancestor of X. laevis and X. borealis, an unknown portion of the early lineage probably evolved in a diploid species; the potential impact of this and other caveats was explored with additional comparative data and analyses (Figs. 1C, D).

We collected expression data from previous studies that used a X. laevis microarray prefabricated by Affymetrix 868788. Expression data was analyzed from five developmental stages or tissue types: egg, embryonic stages 11 and 18, adult testis, and adult ovary. Raw intensity data were converted to CEL files using GeneChip Operation System software (GCOS v. 1.4 Affymetrix). The robust multi array average (RMA) algorithm was implemented to quantify gene expression in GeneSpring version GX7.3 (Agilent, Inc) using either the Affymetrix library file or custom CDF files ("probe masks") that were generated following Hammond et al. 89. The data were then normalized to the median of each gene across all arrays and the 50th percentile of each array. A high intra-treatment correlation (R² = 92–98%) was found between the biological replicates for each treatment.

The Affymetrix X. laevis microarray consists of "probe sets" that are composed of 16 "probe pairs", each of which includes a 25 base pair oligo that is intended to perfectly match the target sequence. Cross hybridization of paralogs could homogenize their expression profiles if it is bidirectional or could amplify differences between them if it is unidirectional. To explore this possibility, we performed a power analysis in which we used probe masks to evaluate paralog specificity of each probe set – i.e. the degree to which the probes on the microarray match one paralog but not the other. We tested three paralog specificities: "low", "medium", and "high". The low paralog specificity analysis included probes that exactly matched (and cross-hybridize to) both paralogs. The medium paralog specificity analysis excluded probes that exactly matched both paralogs. The high paralog specificity analysis excluded probes that perfectly matched both paralogs and also those that had up to and including three mismatches with the non-target paralog. We required each probe set in our analysis to have a minimum of at least 8 probe pairs (and up to 16) at the highest specificity. These probe masks were developed based on comparisons of the probe sequences to a sequences of expressed paralog pairs from previous publications 324142 that were carried out using BLAST searches 80. We evaluated each of these probe specificities under two thresholds for calling presence/absence of expression ("standard" and "conservative" thresholds; Fig. 3). For the "standard" threshold, a paralog was scored as expressed if its raw intensity was above a background level of 50. For the "conservative" threshold, a paralog was scored as expressed if its raw intensity was above a background level of 200. We note that these thresholds are somewhat arbitrary because some probesets may hybridize with lower affinities than others, and therefore recover lower than background raw intensities even though a transcript is in fact expressed. This approach therefore provides only a rough metric of whether or not a transcript is expressed. The Pearson correlation coefficients provide an alternative extreme, because they are based on the expression intensities in all treatments (even those that have below-background raw intensities). These correlation coefficients therefore must be interpreted with the caveat that higher correlations between tissue profiles could be obtained when neither transcript is expressed in many treatments. To contextualize the paralogous correlations, we also calculated the Pearson correlation coefficients between all non-parlaogous expression profiles as in 64.