Introduction

The efficiency of selection against deleterious mutations depends on the selection coefficient and effective population size. As effective population size increases, selection becomes more efficient in purging deleterious mutations (Ohta 1973). Because mammals have small effective population sizes, synonymous sites were thought for many years to evolve neutrally (Graul and Sadee 1997). The question of whether synonymous sites are under selection is important because the ratio of nonsynonymous to synonymous substitution (\(d_{\text{N}} /d_{\text{S}}\)) is commonly used to infer selection in protein-coding genes, and this test assumes that synonymous sites evolve neutrally. If synonymous sites are under selection, then using the \(d_{\text{N}} /d_{\text{S}}\) ratio may overestimate positive selection and underestimate purifying selection. Furthermore, under the assumption of neutrality, synonymous sites are used to compare rates of mutation within genomes and differences in mutation rates among species (Kumar 2005; Kumar and Subramanian 2002; Subramanian and Kumar 2003).

The assumption of neutral evolution at synonymous sites has been challenged by many studies that identified signals of selection at these sites (Bustamante et al. 2002; Doherty and McInerney 2013; Eory et al. 2010; Hellmann et al. 2003b; Ophir et al. 1999; Pagani and Baralle 2004; Smith and Hurst 1998; The Chimpanzee Sequencing and Analysis Consortium 2005). The suggested functional constraints on synonymous sites are (1) signals for the splicing machinery (Caceres and Hurst 2013; Fairbrother et al. 2004; Parmley et al. 2006; Romiguier et al. 2013; Yuan et al. 2012) (2) maintaining mRNA stability (Capon et al. 2004; Duan et al. 2003; Green 2007) (3) binding sites for miRNAs and transcription factors (Gu et al. 2012; Hurst 2006; Stergachis et al. 2013), and (4) efficiency and accuracy of translation (Akashi and Eyre-Walker 1998; Capon et al. 2004; Drummond and Wilke 2008; Duret 2002; Ikemura 1985; Stoletzki and Eyre-Walker 2007; Wright et al. 2004).

Codon usage bias has been primarily observed in species with large effective population sizes such as Drosophila melanogaster, Saccharomyces cerevisiae, Arabidopsis thaliana, and Caenorhabditis elegans (Akashi and Eyre-Walker 1998; Lawrie et al. 2013; Wright et al. 2004; Zhou et al. 2010). In mammals, the correlation between codon bias and tRNA abundance ranges from very weak (Chamary et al. 2006; Doherty and McInerney 2013; Lander et al. 2001) to nonexistent (dos Reis et al. 2004; Duret 2002; Kanaya et al. 2001). Studies using evolutionary comparisons between the divergence at synonymous sites and that at regions that are assumed to be nonfunctional and hence evolve neutrally (e.g., introns, ancestral repeats, and processed pseudogenes) (Berglund et al. 2009; Bustamante et al. 2002; Eory et al. 2010; Hellmann et al. 2003b; Ophir et al. 1999; Smith and Hurst 1998) identify a range of 11–39 % synonymous sites as being under selection. Approximately 1–9 % of synonymous sites at exon–intron boundaries are more conserved than synonymous sites in other regions of the coding sequence and have been proposed to function as exon splice enhancers (Caceres and Hurst 2013; Parmley et al. 2006). Parts of exons have also been identified as exon splice silencers (Wang et al. 2004) and have been shown to be conserved between human and mouse (Wang et al. 2006).

Recent studies that have compared selective constraint between primates and rodents have found that selective constraint is ~2-fold higher in primates than in rodents (Eory et al. 2010; Keightley et al. 2011). Specifically, the percent of synonymous sites under selective constraint in primates and rodents was ~20 and 10 %, respectively. This is contrary to the theoretical expectation, in which the efficiency of selection is expected to be higher in species with large long-term effective population sizes than in species with low effective population sizes. Rodents are estimated to have effective population sizes that are between 2- and 10-fold higher than primates (Charlesworth 2009; Phifer-Rixey et al. 2012). In accordance to theoretical expectations, rodents exhibit a lower \(d_{\text{N}} /d_{\text{S}}\) than either primates or any other mammalian taxon whose genome has been sequenced (Nikolaev et al. 2007; Popadin et al. 2013). To explain the higher selective constraint on synonymous sites in primates than rodents, some authors suggested that the number of exon splice enhancers or exon splice silencers might be higher in primates than in rodents, or that such sequences are under higher selection in primates to maintain functionality over longer life spans (Eory et al. 2010).

To effectively estimate selection on synonymous sites one must control for the effects of neutral processes on the rate of substitution. Some factors affecting the neutral rate of substitution are (1) regional GC content (Hellmann et al. 2005; Tyekucheva et al. 2008; Wolfe et al. 1989), CpG sites (Hwang and Green 2004; Keightley et al. 2011; Siepel and Haussler 2004) (2) DNA replication time (Chen et al. 2010; Pink and Hurst 2010; Stamatoyannopoulos et al. 2009), and (3) recombination rate (Duret and Arndt 2008; Hellmann et al. 2003a; Tyekucheva et al. 2008). Methylated CpG sites in mammals are highly mutagenic because methylated cytosine is unstable and undergoes deamination to thymine (Coulondre et al. 1978). The higher rate of substitution in regions of high GC content can been attributed not only to G and C nucleotides being more mutable than A and T nucleotides (Gojobori et al. 1982) but also to the higher fixation of G and C nucleotides due to biased gene conversion (Duret and Arndt 2008; Meunier and Duret 2004; Ratnakumar et al. 2010). GC-biased gene conversion is a recombination-associated process that results in the biased fixation of G and C nucleotides; as a result hotspots of meiotic recombination have higher rates of substitution than the genomic average. Finally, it has been shown that DNA that replicates late during the S phase undergoes a higher rate of replication-dependent mutation than early replicating DNA (Stamatoyannopoulos et al. 2009). To an extent, these factors have been found to correlate with each other; recombination rate has been found to positively correlate with GC content and to explain a large amount (~47 %) of the variation in GC content (Duret and Arndt 2008). DNA replication time has also been found to vary with GC content, with regions of high GC content replicating early and regions of lower GC content replicating late (Koren et al. 2012; Rhind and Gilbert 2013; Schmegner et al. 2007; Woodfine et al. 2004).

In this study, we estimate selection on synonymous sites using processed pseudogenes as models of strictly neutral evolution. Processed pseudogenes originate through the reverse transcription of mRNAs and their random insertion in to the genome. These sequences are considered “dead on arrival” and are thought to evolve under strict neutrality (Ophir et al. 1999). The use of processed pseudogenes in estimating selection on synonymous sites offers some advantages over the use of other sequences such as introns, intergenic regions, or ancestral repeats. Differences in rates due to base composition are largely constrained since genes and processed pseudogenes share similar sequences. Furthermore, the number of synonymous sites in coding sequences and pseudosynonymous sites in processed pseudogenes is very similar therefore estimates of substitution rate share a similar amount of sampling error.

After assembling a set of orthologous genes and processed pseudogenes in primates and rodents we compared rates of evolution at synonymous sites \((d_{{{\text{S}}_{f} }} )\), at pseudosynonymous sites in pseudogenes \((d_{{{\text{S}}_{\psi } }} )\), as well as the ratio \(d_{{{\text{S}}_{f} }} /d_{{{\text{S}}_{\psi } }}\). If synonymous sites are under selective constraint \(d_{{{\text{S}}_{f} }} /d_{{{\text{S}}_{\psi } }}\) is expected to be significantly lower than one. Because shifts in GC content at synonymous sites in genes or pseudogenes can lead to an increase or decrease in the ratio \(d_{{{\text{S}}_{f} }} /d_{{{\text{S}}_{\psi } }}\), we used a maximum parsimony approach and estimated selective constraint using the six bidirectional mutation patterns (AT ↔ TA, CG ↔ GC, GT ↔ TG, CT ↔ TC, GA ↔ AG, and CA ↔ AC) at each codon site and at fourfold degenerate sites. This allowed us to estimate selective constraint using patterns of mutation that do not affect GC content (AT ↔ TA and CG ↔ GC) versus patterns that can change GC content (GT ↔ TG, CT ↔ TC, GA ↔ AG, and CA ↔ AC).

Materials and Methods

Identification of Processed Pseudogenes and Their Parent Genes

The method used to identify processed pseudogenes is outlined in Section I of the Supplementary Material. Processed pseudogenes from human were used to identify orthologs in chimpanzee (Pan troglodytes), orangutan (Pongo abelii), Rhesus macaque (Macaca mulatta), and common marmoset (Callithrix jacchus). Processed pseudogenes from mouse were used to identify orthologs in rat (Rattus norvegicus). The procedure is explained in Section II of the Supplementary Material.

The parent genes of human processed pseudogenes were used to identify orthologs in chimpanzee, orangutan, macaque, and marmoset, while, mouse parent genes were used to identify orthologs in rat. Orthologs were downloaded from ENSEMBL 72 (Flicek et al. 2013). Parent coding sequences (CDSs) were downloaded for the majority of orthologous processed pseudogenes (4118 human-chimpanzee, 3723 human-orangutan, 2606 human-macaque, 1075 human-marmoset, 1606 mouse-rat). In a few cases there were more than one possible candidate parents (107 human-chimp, 15 human-orangutan, 351 mouse-rat). To identify the parent CDS in such cases, the processed pseudogenes were aligned with their possible parent CDSs using MAFFT (Katoh et al. 2002). Poorly aligned nucleotides were masked using the evaluation mode in T-COFFEE (Notredame et al. 2000) and a CORE score of 5 (Notredame and Abergel 2003). (For more details of the method, see section “Alignment of Processed Pseudogenes and Parent Genes by Codon Positions.”) After alignment refinement the percent similarity was estimated between the processed pseudogene and each possible parent. The CDS with the highest similarity was chosen to be the parent. Because orthologous relations are continuously updated, we further filtered the data according to the most recent version of ENSEMBL (ENSEMBL 77). After alignment refinement, we used the reading frame of the functional coding sequences to determine the positions of homologous pseudocodons in the processed pseudogenes. Aligned codons were discarded if they included gaps, poorly aligned nucleotides, or included stop codons.

Model of Gene and Pseudogene Evolution

We estimated the rate of evolution at synonymous and nonsynonymous sites in genes (\(d_{{{\text{N}}_{f} }} , d_{{{\text{S}}_{f} }} )\) and processed pseudogenes (\(d_{{{\text{N}}_{\psi } }} , d_{{{\text{S}}_{\psi } }} )\) using a maximum likelihood codon model (CODEML) developed by Nielsen and Yang (1998) and implemented in the PAML phylogenetic analysis package (Yang 1997). Specifically, we used the free ratios model to estimate divergences along each lineage (Fig. 1). The free ratios model was implemented using a fixed transition/transversion of 4 (κ = 4) and allowing the κ ratio to be estimated. Equilibrium codon frequencies were estimated using the base composition frequencies at the three codon positions (CodonFreq = 2). To reduce the amount of random error in our estimates, we only used orthologous sets where the number of synonymous sites between orthologous genes and pseudogenes was at least one hundred. This resulted in 1246 estimates from human-chimpanzee comparisons, 950 in human-orangutan, 667 in human-macaque, 566 in human-marmoset, and 392 in mouse-rat. Given that a protein-coding gene may give rise to more than one pseudogene and such cases may bias the data set, a single pair of orthologous genes and corresponding orthologous pseudogenes was chosen. This reduced the data set to 664 orthologous sets in the human-chimpanzee comparison, 547 in human-orangutan, 423 in human-macaque, 427 in human-marmoset, and 217 in mouse-rat. Orthologous genes and pseudogenes were removed if the genes or pseudogenes resided on the X-chromosomes as the X-chromosome has been shown to evolve slower than autosomes (Vicoso and Charlesworth 2006).

Fig. 1
figure 1

In the above phylogeny \(f_{1}\) and \(f_{2}\) depict the codon sequences of a pair of orthologous genes, while \({\psi}_{1}\) and \({\psi}_{2}\) represent the codon sequences of the corresponding pseudogenes. Subscripts 1 and 2 indicate two different species

Full size image

Effects of GC Content on Rate of Synonymous Substitution

To estimate GC content at fourfold degenerate sites (Fig. 1), we only used orthologous genes and pseudogenes that shared at least 40 fourfold degenerate sites. The GC content of fourfold degenerate sites in genes is designated as \({\text{GC}}_{{4{\text{D}}_{f} }}\). For the same set of genes and pseudogenes, GC contents was estimated for each codon site in genes and corresponding “codon” sites in pseudogenes. GC content at these sites is depicted as \({\text{GC}}_{{{\text{codon}}i}}\) were i is the codon position.

To estimate the GC content of sequences flanking the transcriptional start and end sites of genes, we retrieved 5000 nucleotides upstream and downstream of the transcription start site and end site. The notations for GC content at upstream and downstream sequences will be \({\text{GC}}_{{{\text{flank}}_{{f_{\text{up}} }} }}\) \({\text{GC}}_{{{\text{flank}}_{{f_{\text{down}} }} }}\) for genes.

To examine the effects of GC content on \(d_{{{\text{S}}_{f} }},\) we used estimates of GC content from three different regions: (1) fourfold degenerate sites in genes \(({\text{GC}}_{{4{\text{D}}_{f} }} )\) (2) sequences flanking the transcriptional start sites of genes \(({\text{GC}}_{{{\text{flank}}f_{\text{up}} }} )\); and (3) the sequences flanking transcriptional end sites of genes \({\text{GC}}_{{{\text{flank}}_{{f_{\text{down}} }} }}\). Using the lowess function implemented in R we performed regression analyses of \({\text{GC}}_{{4{\text{D}}_{f} }}\), \(GC_{{{\text{flank}}f_{\text{up}} }}\) and \({\text{GC}}_{{{\text{flank}}f_{\text{down}} }}\) against \(d_{{{\text{S}}_{f} }}\). lowess is a locally weighted regression. Because of the limited number of genes in our set, we performed the above analysis using a set of 3059 genes in 13 mammals downloaded from the OMA browser (http://omabrowser.org/; Schneider et al. 2007). The phylogeny (in Newick format) used to estimate \(d_{{{\text{N}}_{f} }}\) and \(d_{{{\text{S}}_{f} }}\) was: (elephant(dog((cow,pig),horse))((rabbit(mouse,rat))(marmoset(macaque(orangutan(chimp,human))))));

Evolution of Pseudononsynonymous and Pseudosynonymous Sites in Pseudogenes

Under the assumption that processed pseudogenes evolve under strict neutrality, the rate of substitution at pseudononsynonymous sites (\(d_{{{\text{N}}_{\psi } }}\)) and pseudosynonymous sites (\(d_{{{\text{S}}_{\psi } }}\)) should be approximately equal. For each species, we compared the log likelihood of the model in which \(d_{{{\text{N}}_{\psi } }} /d_{{{\text{S}}_{\psi } }}\) could be less or equal to 1 to the log-likelihood of the model in which \(d_{{{\text{N}}_{\psi } }} /d_{{{\text{S}}_{\psi } }}\) was equal to 1. Assuming that the former model represents the alternative hypothesis (H1) and the latter model represents the null hypothesis (H0), we can test if \(d_{{{\text{N}}_{\psi } }} /d_{{{\text{S}}_{\psi } }}\) is significantly different from 1 by considering two times the difference between the log-likelihoods of the two models Ln1–Ln0 to be asymptotically distributed as a χ 2 random variable with one degree of freedom.

To further test if pseudogenes are under any selective constraint, we compared mean divergence at pseudononsynonymous (\(\overline{{d_{{{\text{N}}_{\psi } }} }}\)) and pseudosynonymous (\(\overline{{d_{{{\text{S}}_{\psi } }} }}\)) sites using the ratio \(\overline{{d_{{{\text{N}}_{\psi } }} }} /\overline{{d_{{{\text{S}}_{\psi } }} }}\). 95 % confidence intervals were estimated using a bootstrap approach (For further details, see “Bootstrap Analysis” section). To study the effects of the initial GC content at synonymous and nonsynonymous sites on \(d_{{{\text{N}}_{\psi } }}\) and \(d_{{{\text{S}}_{\psi } }},\) we used the GC content at the second codon position (\({\text{GC}}_{{{\text{codon}}2}}\)) as a proxy of the GC content at nonsynonymous sites and the GC content at fourfold degenerate sites (\({\text{GC}}_{{4{\text{D}}_{f} }}\)) as a proxy of the GC content at synonymous sites.

Estimating Selective Constraint on Synonymous and Nonsynonymous Sites

To estimate selection on synonymous and nonsynonymous sites we used two methods: (1) we estimated selective constraint on synonymous sites using the ratio \(\overline{{d_{{{\text{S}}_{f} }} }} /\overline{{d_{{{\text{S}}_{\psi } }} }}\) and selective constraint on nonsynonymous sites using the ratios \(\overline{{d_{{{\text{N}}_{f} }} }} /\overline{{d_{{{\text{S}}_{\psi } }} }}\) and \(\overline{{d_{{{\text{N}}_{f} }} }} /\overline{{d_{{{\text{S}}_{f} }} }}\). By using a bootstrap method, 95 % confidence intervals were estimated. Furthermore, using the lowess function in R we studied the change in \(d_{{{\text{S}}_{f} }}\) and \(d_{{{\text{S}}_{\psi } }}\) with \({\text{GC}}_{{4{\text{D}}_{f} }}\); (2) Using a maximum parsimony approach we estimated the number of substitutions along the gene and pseudogene lineages (Fig. 1). After concatenating the gene-pseudogene codon alignments, we counted the number of substitutions for each codon site and fourfold degenerate sites. Substitutions were separated into the six bidirectional mutation patterns (AT ↔ TA, CG ↔ GC, GT ↔ TG, CT ↔ TC, GA ↔ AG, and CA ↔ AC). To test whether the number of changes along the gene (\(f\)) and pseudogene lineages (\(\psi\)) were significantly different we used a one tail Fisher’s exact test. In instances where the probability was less than 0.05, we used the proportion (\(f/{\psi}\)) as a proxy to selective constraint.

Bootstrap Analysis

To estimate 95 % confidence intervals for the ratios \(\overline{{d_{{{\text{N}}_{\psi } }} }} /\overline{{d_{{{\text{S}}_{\psi } }} }}\), \(\overline{{d_{{{\text{S}}_{f} }} }} /\overline{{d_{{{\text{S}}_{\psi } }} }}\),\(\overline{{d_{{{\text{N}}_{f} }} }} /\overline{{d_{{{\text{S}}_{\psi } }} }},\) and \(\overline{{d_{{{\text{N}}_{f} }} }} /\overline{{d_{{{\text{S}}_{f} }} }}\) we used 10,000 bootstrap samples. Each bootstrap replicate is a random sample of the numerator and denominator of the same size as the original data set.

Results

Alignment Refinement and Estimating Selection at Synonymous Sites

To test the effect of alignment refinement on estimating \(\overline{{d_{{{\text{S}}_{f} }} }} /\overline{{d_{{{\text{S}}_{\psi } }} }},\) we compared the results when using all sites and after removing codons sites masked by different CORE scores. Although there were slight decreases, (<5 %) in both \(d_{{{\text{S}}_{f} }}\) and \(d_{{{\text{S}}_{\psi } }}\), there were no significant changes in the ratios (Fig. 2).

Fig. 2
figure 2

\(\overline{{d_{{{\text{S}}_{f} }} }} /\overline{{d_{{{\text{S}}_{\psi } }} }}\) and 95 % bootstrap CI’s when using different levels of stringency during alignment refinement \(\overline{{d_{{{\text{S}}_{f} }} }} /\overline{{d_{{{\text{S}}_{\psi } }} }}\) does not change by increasing the CORE score

Full size image

Do Processed Pseudogenes Evolve Neutrally?

Processed pseudogenes are formed through the reverse transcription of an mRNA and are randomly inserted in the genome. Because genic regions are found within regions of higher GC content than the genome average, processed pseudogenes are expected to experience a shift in GC content after their formation. As shown in Table 1, the first and third codon positions that include synonymous sites have the highest GC content since they are under weaker selection and are more likely to be influenced by regional GC content. After pseudogene formation, these sites experience the highest shift in GC content. On the other hand, the GC content at the second codon position does not experience a major shift. This is expected if we assume that equilibrium GC content is close to the mammalian average of ~40 % (Arndt et al. 2003; Duret and Arndt 2008).

Table 1 GC content at each codon position (1–3) and fourfold degenerate sites (4D) in genes (f) and pseudogenes (ψ)
Full size table

Assuming that the GC content at second codon sites (\({\text{GC}}_{{{\text{codon}}2}}\)) represents the GC content at nonsynonymous sites and the GC content at fourfold degenerate sites (\({\text{GC}}_{{4{\text{D}}f}}\)) the GC content at synonymous sites, Fig. 3 shows how \(d_{{{\text{N}}_{\psi } }}\) and \(d_{{{\text{S}}_{\psi } }}\) change with GC content. As expected, because the initial GC content at fourfold degenerate sites is higher than the GC content at the second codon sites \(d_{{{\text{S}}_{\psi } }} > d_{{{\text{N}}_{\psi } }}\) when GC content is high. This may occur because processed pseudogenes experience a mutational pressure to lower their GC content after formation. The difference in \(d_{{{\text{S}}_{\psi } }}\) and \(d_{{{\text{N}}_{\psi } }}\) leads to a ratio \(\overline{{d_{{{\text{N}}_{\psi } }} }} /\overline{{d_{{{\text{S}}_{\psi } }} }}\) that is significantly lower than 1 (Table 2 “All”). However when using genes with \({\text{GC}}_{{{\text{codon}}2}}\) and \({\text{GC}}_{{4{\text{D}}f}}\) between 0.3 and 0.5, thus limiting nonequilibrium at pseudosynonymous sites in pseudogenes, \(\overline{{d_{{{\text{N}}_{\psi } }} }} /\overline{{d_{{{\text{S}}_{\psi } }} }} \approx 1\) (Table 2). In conclusion, after controlling for the effect of GC on \(d_{{{\text{S}}_{\psi } }}\), out results indicate that the processed pseudogenes evolve under strict neutrality.

Fig. 3
figure 3

Lowess curves of \({\text{GC}}_{{{\text{codon}}2}}\) against \(d_{{{\text{N}}_{\psi } }}\), and \({\text{GC}}_{{4{\text{D}}_{f} }}\) against \(d_{{{\text{S}}_{\psi } }}\) (shaded areas represent the 95 % confidence intervals). As \({\text{GC}}_{{4{\text{D}}_{f} }}\) content increases, \(d_{{{\text{S}}_{\psi } }}\) increases, and becomes greater than \(d_{{{\text{N}}_{\psi } }}\)

Full size image
Table 2 Testing for evidence of selective constraint in processed pseudogenes. When using all processed pseudogenes, in certain species the assumption of neutral evolution of processed pseudogenes is refuted (\(\overline{{d_{{{\text{N}}_{{{\psi }}} }} }} /\overline{{d_{{{\text{S}}_{\psi } }} }} < 1\)). After controlling for differences in GC content at nonsynonymous and synonymous sites using the GC content at fourfold degenerate sites (\({\text{GC}}_{{4{\text{D}}_{f} }}\)) and the second codon position (\({\text{GC}}_{{{\text{codon}}2}}\)), \(\overline{{d_{{{\text{N}}_{\psi } }} }} /\overline{{d_{{{\text{S}}_{\psi } }} }}\) becomes ~1 (0.3 < GC < 0.5)
Full size table

The Relationship Between GC Content and the Rate of Synonymous Substitution

A positive relation between GC content and the rate of substitution at synonymous sites has been previously identified (Bielawski et al. 2000; Eory et al. 2010; Hurst and Williams 2000). To study how the rate of substitution at synonymous sites (\(d_{{{\text{S}}_{f} }}\)) covaries with GC content, we used the GC content at fourfold degenerate sites (\({\text{GC}}_{{4{\text{D}}_{f} }}\)) and the GC content upstream and downstream of the transcriptional start and end site, respectively \({\text{GC}}_{{{\text{flank}}f_{\text{up}} }}\), \({\text{GC}}_{{{\text{flank}}f_{\text{down}} }}\)) (Fig. 4). As shown in Fig. 4, when \({\text{GC}}_{{4{\text{D}}_{f} }}\), \({\text{GC}}_{{{\text{flank}}f_{\text{up}} }}\) and \({\text{GC}}_{{{\text{flank}}f_{\text{down}} }}\) are at ~0.40 \(d_{{{\text{S}}_{f} }}\) is approximately the same (i.e., regression lines intersect). However, as GC content increases, the trajectories of \(d_{{{\text{S}}_{f} }}\) start to differ (Fig. 4). \(d_{{{\text{S}}_{f} }}\) experiences a sharper increase when using the GC content of sequences upstream or downstream of the transcriptional start and end site of genes. This relation is further verified by a much larger set of genes (Figure S3 in the Supplementary Material).

Fig. 4
figure 4

Variation in the rate of substitution at synonymous sites with GC content at fourfold degenerate sites (\({\text{GC}}_{{4{\text{D}}_{f} }}\), and GC content upstream and downstream of the transcriptional start site (\({\text{GC}}_{{{\text{flank}}_{{f{\text{up}}}} }}\), \({\text{GC}}_{{{\text{flank}}_{{f{\text{down}}}} }}\)). When \({\text{GC}}_{{4{\text{D}}_{f} }}\), \({\text{GC}}_{{{\text{flank}}_{{f{\text{up}}}} }}\), \({\text{GC}}_{{{\text{flank}}_{{f{\text{down}}}} }}\) is ~0.40 \(d_{{{\text{S}}_{f} }}\) is approximately the same, however as GC content at these the location increases the trajectory of \(d_{{{\text{S}}_{f} }}\) start to differ. This is confirmed by a much larger set of mammalian genes (Supplementary Figure S3)

Full size image

Do Synonymous Sites Evolve Under Strict Neutrality?

To estimate selective constraint on synonymous sites we used the ratio \(\overline{{d_{{{\text{S}}_{\text{f}} }} }} /\overline{{d_{{{\text{S}}_{\psi } }} }}\). Under the assumption of strict neutrality and similar mutation patterns at synonymous and pseudosynonymous sites in pseudogenes (i.e., when both sites are evolving under equilibrium conditions), \(d_{{{\text{S}}_{f} }}\) should evolve at approximately the same rate as \(d_{{{\text{S}}_{\psi } }}\). As shown in Fig. 4, \(d_{{{\text{S}}_{f} }}\) increases when \({\text{GC}}_{{4{\text{D}}_{f} }}\), \({\text{GC}}_{{{\text{flank}}f_{\text{up}} }},\) or \({\text{GC}}_{{{\text{flank}}f_{\text{down}} }}\) are greater than ~0.4. If we assume that the value of \(d_{{{\text{S}}_{f} }}\) is similar to what is expected under equilibrium and if we further assume that nonequilibrium evolution at pseudosynonymous sites in pseudogenes is limited when \({\text{GC}}_{{4{\text{D}}_{f} }}\) and \({\text{GC}}_{{{\text{codon}}2}}\) is ~0.40 (Fig. 3), then any deviation of \(d_{{{\text{S}}_{f} }} < d_{{{\text{S}}_{\psi } }}\) from 1 could be the result of selective constraint.

As shown in Fig. 5, when \({\text{GC}}_{{4{\text{D}}_{f} }} \approx 0.4\) in most primates \(d_{{{\text{S}}_{f} }} < d_{{{\text{S}}_{\psi } }}\); while in rodents \(d_{{{\text{S}}_{f} }} \approx d_{{{\text{S}}_{\psi } }}\). This is further demonstrated by the ratios \(\overline{{d_{{{\text{S}}_{f} }} }} /\overline{{d_{{{\text{S}}_{\psi } }} }}\) in Table 3. Contrary to the higher selective constraint on nonsynonymous sites exhibited by rodents (\(\overline{{d_{{{\text{N}}_{f} }} }} /\overline{{d_{{{\text{S}}_{\psi } }} }}\) or \(\overline{{d_{{{\text{N}}_{f} }} }} /\overline{{d_{{{\text{S}}_{f} }} }}\)), primates exhibit a higher selective constraint on synonymous sites. This is contrary to the theoretical expectation in which the efficiency of selection against deleterious mutations is higher in species with larger long-term effective population sizes than species smaller long-term effective population sizes (Ohta 1973).

Fig. 5
figure 5

Variation of \(d_{{{\text{S}}_{f} }}\) and \(d_{{{\text{S}}_{\psi } }}\) with \({\text{GC}}_{{4{\text{D}}_{f} }}\). When \({\text{GC}}_{{4{\text{D}}_{f} }}\) content is ~0.40, \(d_{{{\text{S}}_{\psi } }}\)  >  \(d_{{{\text{S}}_{f} }}\) in all primates except chimpanzee; while in rodents \(d_{{{\text{S}}_{f} }} \approx d_{{{\text{S}}_{\psi } }}\).When \({\text{GC}}_{{4{\text{D}}_{f} }}\) content increases, in all species except rat \(d_{{{\text{S}}_{f} }}\) can become larger than \(d_{{{\text{S}}_{\psi } }}\). This change in rate between \(d_{{{\text{S}}_{f} }}\) and \(d_{{{\text{S}}_{\psi } }}\) is most prominent along the mouse lineage

Full size image
Table 3 Testing selective constraint on synonymous and nonsynonymous sites. Primates show evidence of selection on synonymous sites (\(\overline{{d_{{{\text{S}}_{f} }} }} /\overline{{d_{{{\text{S}}_{\psi } }} }} < 1\)) while in rodents synonymous sites evolve neutrally (\(\overline{{d_{{{\text{S}}_{f} }} }} /\overline{{d_{{{\text{S}}_{\psi } }} }} \approx 1\)). On the contrary, rodents show significantly higher selective constraint on synonymous sites (\(\overline{{d_{{{\text{N}}_{f} }} }} /\overline{{d_{{{\text{S}}_{\psi } }} }}\) and \(\overline{{d_{{{\text{N}}_{f} }} }} /\overline{{d_{{{\text{S}}_{f} }} }}\))
Full size table

A possible contribution to the larger deviation from neutrality in primates could be nonequilibrium evolution at pseudosynonymous sites in pseudogenes. To further control for such effects, we used a maximum parsimony approach and estimated selection on nonsynonymous and synonymous sites in genes using six bidirectional mutational patterns. This would allow us to study how selective constraint is affected by mutational patterns that lead to a shift in GC content (GT:TG, CT:TC, GA:AG, CA:AC) and ones that do not (AT:TA, CG:GC). Selective constraint was estimated for each codon position and fourfold degenerate sites. As shown in Table 4, rodents show higher selective constraint on the first and second position all across the different mutational patterns, while at the third codon position selective constraint becomes similar. As expected, on average, selective constraint on the third codon position is the lowest when using GA or CT changes because most of them code for the same amino acid. When using fourfold degenerate sites almost all mutational patterns do not show a significant difference between genes and pseudogenes. However, in the case of macaque and marmoset, we see some significant deviations in the case of CA and CT mutations (Table 4). These can be caused by pseudogenes experiencing a mutational pressure to lower their GC content. Because of the large sample of changes we looked at unidirectional changes (C → A, A → C, C → T, T → C) in macaque and marmoset genes (\(f\)) and pseudogenes (\(\psi\)). As shown in Table 5, only C → A and C → T mutations show a significant difference between genes and pseudogenes. This supports our hypothesis that the significant deviations in the rates of synonymous substitutions between genes and pseudogenes observed in macaque and marmoset (Fig. 5; Table 4) are most likely caused by pseudogenes experiencing nonequilibrium to lower their GC content. The significant difference in total changes observed in orangutan can be caused by the cumulative effect of mutational patterns that are close to being significant. As in the case of macaque and marmoset, the largest differences are observed by CA and CT changes.

Table 4 Estimating selective constraint on each codon site and fourfold degenerate sites using bidirectional mutational patterns
Full size table
Table 5 A breakdown of macaque and marmoset bidirectional mutational patterns showing significance at fourfold degenerate sites (Table 4)
Full size table

Discussion

A frequently used method of estimating selection at synonymous sites in genes is comparing the rate of substitution at regions assumed to be nonfunctional and under no selection, to the rate at synonymous sites (Bustamante et al. 2002; Eory et al. 2010; Hellmann et al. 2003b; Ophir et al. 1999; Subramanian and Kumar 2003). Because the frequency of insertions and deletions in nonfunctional regions is much higher than in protein-coding genes, alignment errors can have a significant impact on the estimate of selective constraint on synonymous sites. Using different levels of alignment quality, we show that the quality level does not have a significant impact on \(\overline{{d_{{{\text{S}}_{f} }} }} /\overline{{d_{{{\text{S}}_{\psi } }} }}\) (Fig. 2).

When testing for selective constraint on synonymous sites we assume that pseudosynonymous sites in pseudogenes are under no selective constraint. To test the above assumption, we compared the rate of substitution at pseudononsynonymous and pseudosynonymous sites in pseudogenes \(\overline{{d_{{{\text{N}}_{\psi } }} }} /\overline{{d_{{{\text{S}}_{\psi } }} }}\). Assuming that processed pseudogenes are evolving under no selective constraint \(\overline{{d_{{{\text{N}}_{\psi } }} }} /\overline{{d_{{{\text{S}}_{\psi } }} }} \approx 1\). Although our initial estimates refuted the null hypothesis of processed pseudogenes evolving under strict neutrality (\(\overline{{d_{{{\text{N}}_{\psi } }} }} /\overline{{d_{{{\text{S}}_{\psi } }} }} < 1\)) (Table 2 “All”). However, when controlling for the difference in GC content at synonymous and nonsynonymous sites in pseudogenes \(\overline{{d_{{{\text{N}}_{\psi } }} }} /\overline{{d_{{{\text{S}}_{\psi } }} }}\) becomes approximately equal to 1 (Table 2 “0.3 < GC < 0.5”).

Under the assumption that processed pseudogenes evolve under strict neutrality selective constraint on synonymous sites can be estimated using the ratio \(\overline{{d_{{{\text{S}}_{f} }} }} /\overline{{d_{{{\text{S}}_{\psi } }} }}\). Because selection at synonymous sites is weak, differences in mutation patterns or rates of biased gene conversion between genes and pseudogenes can cause complications in estimating the extent of selective constraint on synonymous sites (Comeron 2006; Lawrie et al. 2011; McVean and Charlesworth 1999). For example, if synonymous sites are located within hot spots of recombination, \(d_{{{\text{S}}_{f} }}\) can be equal or even larger than \(d_{{{\text{S}}_{\psi } }}\), therefore masking any conservation due to purifying selection. To the opposite effect, because processed pseudogenes usually move from regions of high GC content to areas of low GC content they experience a mutational pressure to lower their GC content and therefore this can cause \(d_{{{\text{S}}_{\psi } }}\) to be larger than \(d_{{{\text{S}}_{f} }}\), therefore creating the false impression of selective constraint on synonymous sites.

As shown in Fig. 4 and Figure S3 (Supplementary Material), any deviations from a GC content of ~0.4, \(d_{{{\text{S}}_{f} }}\) experiences a significant increase. Assuming an equilibrium GC content similar to the average GC content of primate and rodent genomes (~0.40) (Arndt et al. 2003; Duret and Arndt 2008), an increase in \(d_{{{\text{S}}_{f} }}\) with GC content could be caused by CpG deamination, biased gene conversion, or an increase in the mutation rate from AT to GC nucleotides (Piganeau et al. 2002). If we assume that the value of \(d_{{{\text{S}}_{f} }}\) when \({\text{GC}}_{{4{\text{D}}_{f} }} \approx 0.4\) is the least affected by the above factors, when comparing it to \(d_{{{\text{S}}_{\psi } }}\) it seems that synonymous sites are evolving at a significantly lower rate in primates, while in rodents, they evolve at the same rate as the neutral expectation (Table 3; Fig. 5). On the contrary, rodents exhibit a higher selective constraint on nonsynonymous sites (Table 3) which is in accordance to the theoretical expectation in which species with larger effective population sizes are more efficient in purging deleterious mutations (Ohta 1973).

To further ensure that the significant difference between \(d_{{{\text{S}}_{f} }}\) and \(d_{{{\text{S}}_{\psi } }}\) in primates is not due to pseudogenes experiencing a higher rate of GC to AT mutations after formation, we estimated selective constraint on each codon position and at fourfold degenerate sites using a maximum parsimony approach. Specifically, we separated mutations along the gene and pseudogene lineages into six patterns (GT:TG, CT:TC, GA:AG, CA:AC, AT:TA, CG:GC). We did not look at unidirectional changes because of the very small number of changes in certain patterns. As shown in Table 4 selective constraint on the first and second position is higher in rodents; but when selective constraint becomes weaker such as the third codon position the difference in selective constraint disappears. When it comes to fourfold degenerate sites in which all mutational changes do not alter the amino acid, almost all mutation patterns indicate no selective constraint on synonymous sites. In the case of macaque and marmoset, however, we observe a very large difference in CA:AC and CT:TC mutations between genes and pseudogenes. If these differences are caused by pseudogenes experiencing a mutational pressure to lower their GC content, we expect to see a significant difference in patterns C → T and C → A. As expected, only patterns C → T and C → A show a significant difference (Table 5). This difference between primates and rodents can be caused by the higher mutability of CpG sites in primates (Keightley et al. 2011).

Our results help resolve previous puzzling findings (Eory et al. 2010; Keightley et al. 2011) in which synonymous sites in rodents are under lower selective constraint than primates, despite having larger effective population sizes (Charlesworth 2009). According to the results of the present study, synonymous sites are under no selective constraint. Although we can conclude with high confidence that the percent of synonymous sites under selective constraint is not significantly greater than 10 %, our method may not be sensitive enough to detect if 10 % or less of synonymous sites are under selection as suggested by human population data (Keightley and Halligan 2011) or as exon spice site studies (Caceres and Hurst 2013; Hurst 2006). Some of the shortcomings of our study is that our sample of genes is limited to ones giving rise to pseudogenes, and furthermore, our study did not control for any effects associated with transcription coupled repair. If transcription-associated processes are mutagenic (Green et al. 2003; Majewski 2003) then \(\overline{{d_{{{\text{S}}_{f} }} }} /\overline{{d_{{{\text{S}}_{\psi } }} }}\) can be overestimated if processed pseudogenes are at large transcriptionally inactive.