Until recently, evolutionary genomics questions, including questions about amino acid composition of proteins, patterns of stabilising and positive selection and mechanisms of retention of duplicated genes and new function evolution, were typically answered either by analyzing phylogenies of select gene families 12 or by full-genome analysis of triplets of genomes with two ingroup genomes compared to measure evolutionary rates, while the third, outgroup, genome used to polarize the observed changes 3. As the strategy of genome sequencing shifts from broad taxonomic coverage to sequencing multiple closely related genomes 4, a need arises in a set of tools to accomplish a phylogenetic analysis of amino acid substitutions in coding portions of a large number of protein families simultaneously and to address the question of generality of patterns observed in limited and possibly biased set of select gene families. Questions that can be asked using such approach include, but certainly are not limited to enquiries about long-term changes in amino acid compositions of proteins, about selective constrains and pressures across the genome and evolution of novel gene functions by retention and modification of duplicated genes. Here we present a tool to accomplish phylogenetic analysis of amino acids substitutions on the whole-genome scale using multiple amino acid alignments of over 11,000 gene families from twelve completely sequenced Drosophila genomes and illustrate its utility by the analysis of the resulting database of amino acid substitutions spanning 70 million years of drosophilid proteins evolution.

Global patterns of amino acid compositions of proteins is thought to not be at a detailed balance, but rather appears to be gradually evolving by consistently adding rare to and removing common amino acids from the amino acid repertoire of protein sequences 35. There is an on-going debate on whether is pattern reflects the order in which amino acids have been added to the genetic code 36 or is caused by biases in mutability of particular codons 789. As pointed out by 8, one way to address this controversy is to analyze the observed trends in a range of genomes of increasing degree of divergence: if the observed patterns are caused by the effect of amino acid polymorphism reflecting mutation-selection biases they are expected to become less pronounced as divergence between genomes increases. Furthermore, there may be substantial differences in selection pressures on reciprocal amino acid substitutions 10: changes from polar to non-polar amino acids in human proteins are more permissive than vice-versa. Such asymmetry and the degree to which is can contribute to the large-scale changes in amino acid composition has not yet been measured on the scale of several genomes.

Differences in patterns of selective pressure have also been predicted between evolutionary retained duplicated genes and single-copy genes 11121314. Duplicated genes can persist in genomes either because one of the copies has acquired a new function (neofunctionalization 151617), or because both copies are now needed to perform the function or functions previously accomplished by a single copy (subfunctionalization). Subfunctionalizaton can occur either by means of partitioning of the ancestral functions between the two copies (for example by loss of one of alternative promoters in each copy), or by means of balanced degradation, i.e., fixation of hypomorphic alleles in each copy 1819. Each of these mechanisms implies relaxation of stabilizing selection, resulting in faster evolution in paralogs than in single-copy genes. Specifically, pure neofunctionalization occurs by accumulation of mutations in one of the copies, while the other remains under stabilizing selection 13141617. Subfunctionalization occurring through balanced degradation, on the other hand, is accompanied by accumulation of deleterious mutations in both paralogs. Finally, subfunctionalization occurring by tissue- or developmental stage-specialization of gene expression without a change in functionality would result in retention of stabilizing selection action in both paralogs. It is much harder to make predictions about other types of subfunctionalization, such as subdivision of pre-existing multiple substrate specificity between duplicated genes, because the two functions may depend on different parts of coding portion of the gene and, therefore, retaining one but not the other may relax selective constraints acting on at least part of the sequence. Previous studies of duplicated genes in Drosophila genomes (e.g., 19) detected elevated signal of positive selection in a subset of gene families with duplications using K_a/K_s approach. Here we report a genome-wide analysis of differences between duplicated and single copy genes in frequency and spectrum of amino acid substitutions.

Several caveats in the data and analysis require attention. Firstly, alignments we used may contain pairs of paralogs, in which one of the copies is undergoing pseudogenization and is nor longer expressed, but has not yet acquired a frameshift, which would allow it to be recognized as a pseudogene. Indeed, there is a significant excess of nonsense mutations (per missense) present in the terminal branches of phylogeny (data not presented), indicating presence of pseudogenes in the alignments. Pairs of paralogs, in which one gene copy is undergoing pseudogenization, will demonstrate clade asymmetry, mimicking the signature of neofunctionalization. However, such paralogs are almost certainly present only in the most terminal branches of Drosophila phylogeny spanning over 70 mln years, because the half-life of duplications, in which one of the copies undergoes pseudogenization, is 2-4 mln years (12; 26). Terminal branches include a minority of duplications in our database and excluding such branches from the analysis does not alter the results (Table 1). This indicates that the observed clade asymmetry is not an artefact of pseudogenes. A direct comparison of clade asymmetries in terminal vs. non-terminal duplications is not possible for two reasons. Firstly, there are much fewer substitutions in the terminal branches, so there is an intrinsic difference in statistical power. Secondly, clade asymmetry analysis is based on unambiguous substitutions and the frequency of unambiguous substitutions increases with the depths of the phylogeny, possibly biasing such comparison.

On the other hand, some true functional paralogs may be missing from the alignments, particularly those resulting from ancient duplications, due to homology below the threshold used by the reciprocal BLAST algorithm (see Methods). This creates a bias towards less divergent paralogs, reducing our ability to detect elevated rates of evolution in duplicated genes. Relative magnitude of these opposing biases remains unknown.

Further, results presented in Table 1 do not necessarily indicate that clade asymmetries are more likely to manifest themselves in substitution rates than in substitution radicality. The number of test surviving multiple test correction probably reflects differences in statistical power rather than a true biological phenomenon.

Systematic loss/gain asymmetry in amino acid composition in 12 Drosophila genomes corroborates patterns previously observed in a variety of taxonomically diverse triplets of genomes 3. This pattern does not become less pronounced as more and more distant genomes are included into consideration, indicating that it is not caused by the effect of polymorphisms reflecting mutation-selection balance influenced by mutational asymmetries 79.

We also demonstrate that this net loss/gain asymmetry is strongly correlated with source and destination amino acid polarities: substitutions of polar amino acids by non-polar ones have a higher net rate than the reciprocal substitutions. In the past we have demonstrated a similar polarity-related asymmetry in selection coefficients against amino acid substitutions in human proteins 10; however this asymmetry was largely limited to strong selection (i.e., selection against clinically important phenotypes) and was not seen in evolutionary substitution rates.

One may hypothesise that replacing polar amino acids by any is less disruptive for the protein function because polar amino acids have a lower tendency to be located internally in the tertiary protein structure 21. If so, we would expect the decrease of polarity due to amino acid substitutions to be the lowest in membrane proteins, in which polar amino acids in within-membrane domains tend to be internally located. Indeed, the decrease of polarity due to substitutions is the weakest in receptor and transporter proteins, many of which have membrane-embedded hydrophobic regions (Fig. 2 B) and in proteins with membrane localization (Fig. 2 C).

A question remains how is it possible that asymmetry in amino acid gains and losses systematically removed polar amino acids more often than non-polar ones (Fig.2A) over 70 mln years of drosophilid evolution (and actually over much longer period of evolution of proteins of much broader taxonomic spectrum 3)? There is no evidence that the relationship shown on Fig. 2A has a tendency to weaken in the most recent branches of the phylogeny (data not reported), which would have indicated an approach to an equilibrium. Rather, the frequencies of amino acids in proteins appear to be far from an equilibrium and we observe a constant turnover of polar amino acids due to more relaxed selective constraint acting on the amino acid of external location. One may further speculate that perhaps such systematic loss of surface polar amino acids would gradually change protein folding as external sites become occupied by more hydrophobic amino acid residuals. This process may be a potentially important mechanism of acquiring new functions by duplicated genes.

We have demonstrated that, in a genome-wide assessment, duplicated genes evolve both faster (higher K_a/K_s) and through more radical amino acid substitutions (higher |DPolarity|, lower exchangeability) than single copy genes (Figs 3 and 4). Likewise, single copy genes in families with extant duplications tend to evolve faster and more radically than single copy genes in families without extant duplications, indicating that duplications are more likely to be retained in gene families with weaker selective constraints.

Just like with the signed polarity change, the absolute change of polarity is not significantly different between duplicated and single copy genes among genes coding for transporter proteins, corroborating the hypothesis of the importance of relaxed selective constraint on surface sites of water-soluble proteins (Fig. 4 A). (This difference is, however, significant for receptor proteins.) The exchangeability index, on the other hand, is significantly lower in duplicated transporter proteins, suggesting that paralogs in these genes families do evolve through more radical substitutions, just without systematic net loss of polar residuals.

Data on the asymmetry of clades resulting from duplications supports the hypothesis of widespread neofunctionalization accompanying retention of duplicated genes: over 1/3 of all duplications show a significant asymmetry in amino acid substitution rates with false discovery rate 0.05 and almost 1/5 of all substitution show asymmetry, which stands Bonferroni correction (Table 1). Much fewer duplications show a significant asymmetry in radicality of substitutions, although about 6% have a significant asymmetry in absolute polarity change (with false discovery rate 0.05). Gene families of different functionality differ from each other in the degree of clade asymmetry with a hint of a negative correlation between asymmetry in rates (Fig. 6, top) and asymmetry in radicality (Fig. 6, middle and bottom). No molecular function category stands out in terms of tendency to display signatures of neofunctionalization, although RNA-binding proteins have the lowest (non significant) difference in rates and radicality of substitutions between duplicated and single copy genes (Fig. 3 and 4) and the lowest clade asymmetry of substitution radicality in paralogs (Fig. 6), indicating that, perhaps, in these proteins neofunctionalization is less common. Interestingly, transcription factors appear to show low neofunctionalization signal in terms of substitution rates (no difference between duplicated and single-copy genes, Fig. 3; low asymmetry between paralogs, Fig 6, top), but a strong neofunctionalization signal in terms of substitution radicality (Fig 4; Fig. 6 middle and bottom). One may hypothesize that positive selection for a novel functionality can operate either by increased rate of substitutions, or by favouring more radical changes without the increase of rates.

We have designed a tool, which allows a detailed phylogenetic analysis of amino acid substitutions in a large number of multiple alignments with or without duplicated genes present. The algorithm is capable to polarize and establish phylogenetic position of all substitutions for which it is possible (unambiguous) and to list all possible alternatives for other, ambiguous substitutions. It results in a database, which can be used to answer questions about patterns of amino acids substitutions genome-wide or in particular categories of genes such as molecular functions or duplication status.

The analysis of such database of substitutions in 12 Drosophila genomes confirmed previously observed non-equilibrium patterns of net losses and gains of individual amino acids, demonstrated that these patterns do not weaken with the depth of phylogeny and revealed a strong correlation between polarity of amino acid and propensity to display a net loss. We hypothesize that this effect can be explained by relaxed selective constraints on externally located amino acid sites occupied by polar residuals. Evolution of duplicated genes is characterized by both higher relative rate of substitution and more radical nature of these substitutions, as compared to single copy genes. The rate and radicality in paralogs displays a weaker relation with mean expression rate and variance of expression rates across tissues than in single copy genes. This pattern, along with the strong asymmetry between clades resulting from duplication events, indicates widespread neofunctionalization of retained duplications.

LYY proposed the study methodology, accomplished data analysis and prepared the manuscript. MAB wrote software, generated the dataset and contributed to the manuscript preparation.

The authors declare that they have no competing interests.