Exons from eight metazoan species (Human (Hs), mouse (Mm), Danio rerio (Dr), Caenorrhabditis elegans (Ce), Caenorrhabditis briggsae (Cb), Anopheles gambiae (Ag), Drosophila melanogaster (Dm), Apis mellifera (Am)), one plant (Arabidopsis thaliana (At)) and two ascomycetous fungi (S. cerevisiae (Sc), Schizosaccharomyces pombe (Sp)), were examined for trends in amino acid composition as one approaches the exon-intron boundary. Species were chosen from among a relatively small set of organisms for which high quality comparative data on splice-regulatory proteins have recently become available 13. As splice site signals can extend into exons and our focus is on exonic splicing regulation, we removed the first full codon at the exon-intron boundary (see Materials and methods). Thereafter, rank correlations (rho) between distance from the boundary (34 codons into the exon; see Materials and methods) and proportional usage of the amino acid were computed independently for 5' and 3' regions of exons. Further, for all amino acids independently we fitted a linear regression extracting the slope of the line to be used as a crude diagnostic for the strength of amino acid preference/avoidance. Figure 1 illustrates the different types of relationship observed.

Two-fold and four-fold blocks of the six-fold degenerate amino acids were considered as distinct groupings so that a total of 46 tests (23 amino acid groups 5' and 3') were carried out for each species. Tables 1 and 2 give a comprehensive by-species overview of amino acid preferences/avoidances, significant after Bonferroni correction (N = 46 comparisons, P < 0.0011). Additional data file 1 contains the complete set of rank correlations for all 11 species.

The most conspicuous feature of Tables 1 and 2 is arguably the commonality of trends in the metazoa and the scarcity of trends in the ascomycetous yeast species. The two-fold block of leucine (L2) in S. cerevisiae is the only amino acid grouping exhibiting a significant preference trend (rho = -0.4482, P < 0.0003). This is in stark contrast to the suite of multicellular eukaryotes where an extensive range of avoidance and preference trends is observed. Only three multicellular species display fewer than 13 significant trends (Dm, Ag, At) whereas five (Hs, Mm, Ce, Cb, Am) display more than 20. For D. melanogaster and C. elegans, we tested whether the results might be biased as a result of exon homology, but in either case found amino acid abundance patterns at exon ends to be virtually identical in a set of homology-reduced genes (Dm, N = 8,840; Ce, N = 11,790; Additional data files 2 and 3).

The role of exonic guidance in splicing organization has been linked to multiple aspects of genome composition and pre-mRNA structure, including intron/exon length 1516, intron number 7 and density 17 and splice site information content 71819. The number of significant amino acid trends per species tightly covaries with some of these factors, notably the mean number of introns per gene (rho = 0.95, P < 0.0001), median coding sequence (median CDS) per gene (rho = -0.97, P < 0.0007), genomic number of introns (rho = 0.86, P < 0.003), and intron length (log10(mean length): rho = 0.83, P < 0.006) as expected under a model where complex transcripts with multiple long introns elicit increasing reliance on exon definition 15. On the other hand, neither SR protein family size (rho = 0.59, P = 0.09) nor splice site information content (5', rho = -0.26, P = 0.50; 3', rho = 0.43, P = 0. 25) show any relationship with the number of amino acid skews near intron-exon boundaries. The latter observation is perhaps the more interesting as it suggests that there is no straightforward compensatory relationship between splice site information content and the need for exonic regulation across species.

Finally, the number of exons from which amino acid trends were derived, although correlated with the number of trends (rho = 0.86, P < 0.003), does not feature among the top predictors when multicollinearity is controlled for (Additional data files 4-6). Together with the observation that we find relatively few trends in Arabidopsis, despite the substantial number of exons sampled, this suggests that sample size is not the critical factor in detecting different numbers of trends across species. We must stress, however, that the above results should be regarded as strictly exploratory given the small number of observations (Additional data file 4). A greater number of species with more comprehensive phylogenetic sampling will be required to validate the results in the future.

The preeminence of exon-intron structure in predicting the number of amino acid trends suggests that the intron-poor ascomycetous fungi analyzed here might not be representative of their kingdom. We therefore analyzed the composition of exon ends in Cryptococcus neoformans (Cn), an intron-rich basidiomycete. Strikingly, we find a large number (26) of preference and avoidance trends in this species (Table 3 and Additional data file 1), with some marked similarities in comparison to metazoan trends, particularly 5'. Furthermore, the inclusion of C. neoformans data in the analysis of potential predictor variables does not substantially change previous results: the mean number of introns per gene (rho = 0.91, P < 0.0002), median CDS per gene (rho = -0.68, P < 0.032) and the genomic number of introns (rho = 0.72, P < 0.02) remain strong predictors (Additional data file 6).

Virtually nothing is known about the splicing mechanism in C. neoformans but the demonstration of alternative splicing pathways in this species 20 as well as low splice site information content (Additional data file 5) 7 make the presence of exonic splicing regulation a credible possibility. Consistent with this, the predicted C. neoformans proteome contains multiple proteins resembling known eukaryotic SR proteins, particularly in that they harbor RNA recognition domains (Additional data file 7). This is suggestive of involvement in splicing, albeit evidently insufficient to reach conclusions about specific functional roles of these proteins.

Whilst the spectra of amino acids preferred/avoided by individual species are ultimately unique in breadth (how many trends) and composition (which amino acids are affected), there is considerable cross-specific overlap in terms of whether a particular trend is present at all, its direction, and relative strength (as measured by the slope of the line of best fit). Tables 1 and 2 illustrate that this particular agreement is virtually perfect between human and mouse 2, with marginal differences in the relative strength of individual trends, and that directionality is conserved throughout. Considering zebrafish (Dr) as the only other vertebrate in our sample alongside these species, we notice that its spectrum is slightly diminished in breadth and contains a few trends not seen in the two mammals (G (3'), V (5',3')). However, overall concordance in composition and strength is still remarkably good, and the 'mammalian pattern of directionality' perfectly adhered to. The nematode pair almost matches the human-mouse dyad in terms of overall concordance of preference patterns, with directionality perfectly conserved.

For the most part, the patterns of preference/avoidance are repeatable across species. Table 4 shows pairwise comparisons between species giving rank correlations (rho) for the slopes derived from all 23 amino acid groupings. For the vertebrate group both 5' and 3' correlations are very high (all rho > 0.9, all P < 1.81E-06; 90 tests, significance threshold, P < 5.56E-04), with human and mouse in almost perfect agreement. More remarkably, however, some strong correlations also exist 3' between the vertebrates and, for example, Anopheles (all rho > 0.87, all P < 2.94E-06) and Drosophila (all rho > 0.75, all P < 2.9E-05). The 3' correlations are less impressive for the remaining species (Am, At, Cn) but Apis (all rho > 0.75, all P < 4.11E-05) and even Cryptococcus (all rho > 0.69, all P < 5.56E-04) boast remarkably strong 5' correlations with the vertebrates. Focusing on specific amino acid trends, isoleucine (I) stands out in that it is strongly preferred near 3' boundaries across all species; others are well represented, albeit not universal, through the entire phylogeny - for example, 5' avoidance of glutamine (Q), and 3' preference for phenylalanine (F).

The strong cross-species concordance in preference patterns makes one observation all the more striking. The nematode 5' spectra behave in a highly counterintuitive manner in that the 'mammalian pattern of directionality' is violated on several occasions: where we do find significant trends in nematodes and other species (E, K, L2, Q, R4, R2, T), all but glutamine (Q) show discrepant directionality (Table 2). For example, whereas lysine (K) is strongly preferred near boundaries in vertebrates and some insects (Dm, Am), it appears to be strongly avoided in the 5' region of nematode exons (Figure 2). Table 4 also underlines the exceptional position of nematodes: 5' correlations between nematodes and any other species are pervasively negative. No single correlation across all amino acids is significantly different from zero applying the adjusted significance threshold (P < 5.56E-04), owing to several trends collapsing into insignificance rather than fully reversing sign. However, the pervasiveness of this pattern is nonetheless noteworthy, especially considering that the same is not the case for the 3' spectra where we find a coherent agreement between nematodes and vertebrates (minimum rho > 0.65, all significant at P < 5.92E-04) and only the two-fold block of serine (S2) shows a reverse pattern of directionality among the significant trends for individual amino acids.

This curious discrepancy between 5' and 3' spectra of amino acid trends in nematodes led us to investigate further the relationship of 5' and 3' patterns across species. Considering all amino acid trends simultaneously, rank correlations between slope coefficients (5'~3') were computed. Furthermore, we wanted to explicitly test the hypothesis that preference trends show a 'symmetric' behavior, that is, that individual amino acids exhibit preference trends of similar strength and direction at 5' and 3' ends. To this end, we carried out standardized major axis regressions (SMA; see Materials and methods) 2122 for 5' versus 3' trends in each species and compared the resulting regression line with one expected under perfect symmetry (y = x). The results are given in Table 5 and graphically represented in Figure 3. Human and mouse show very substantial positive correlations between 5' and 3' preference trends (Hs, rho = 0.8528, P = 1.96E-06; Mm, rho = 0.8626, P = 2.28E-06). Although diminished in strength, we also see significant correlations for Drosophila and Danio. As expected from the previous analysis, correlations for nematodes are negative, albeit not significantly so (Ce, rho = -0.1413, P = 0.5185; Cb, rho = -0.4358, P = 0.0388). However, the SMA results allow us to reject any notion of C. elegans or C. briggsae adhering to a symmetric pattern of amino acid usage, the respective confidence intervals (CIs) ruling out a symmetry slope of β = 1 (CI (Ce), [-1.118; -0.7309]; CI (Cb), [-0.7474; -0.5139]). No other species for which an SMA could be carried out (Table 5; Materials and methods) deviate significantly from a symmetric model, although symmetry of amino acid trends varies greatly and can only really be called a defining characteristic of exon ends in vertebrates.

Intriguingly, asymmetries in the amino acid composition of nematode exon ends appear to be mirrored by a corresponding asymmetry of regulatory motifs. Robinson 23, using a computational approach to characterize candidate ESEs in C. elegans, found that 5' and 3' ends were distinguished by different classes of consensus motifs. Crucially, he found purine-rich human-like candidate motifs to be associated with 3' ends but not 5' ends of nematode exons, which is broadly consistent with our observation that amino acids encoded by purine-rich codons tend to be, in contrast to other animals, disfavored at 5' ends (Table 2 and Figure 3).

For mammals, the prediction that amino acids preferred near boundaries should correspond to those favored in ESEs was tested by Parmley et al. 2. The authors defined a metric that quantifies the involvement of amino acids in splice enhancer hexamers relative to the null expectation that every codon is represented in ESEs around its genomic frequency. As predicted, these hexamer preference indices (HPIs), computed for each amino acid grouping, were found to correlate with preference trends, strongly preferred amino acids on average associated with higher HPI values.

This relationship holds true for human as well as murine ESE sets and amino acid trends, considering either rank correlation coefficients (rho_x; Hs HPI~rho_x, rho = -0.54, P < 0.00001, N = 46; Mm HPI~rho_x, rho = -0.49, P = 0.0005, N = 46) or the slope (β) of the fitted linear model (Hs HPI~β, rho = -0.57, P < 0.0001, N = 46; Mm HPI~β, rho = -0.52, P = 0.0002, N = 46).

As expected from the demonstration that ESEs can act at varying distances from the splice site 14, human ESEs do not exhibit a reading frame bias beyond what is expected from the genomic frequencies of the underlying codons (Additional data file 8). They can also, in principle, incorporate most codons (Additional data file 8). In consequence, the defined set of amino acids we find avoided or preferred are likely not due to ultimate exclusion of certain codons but because different efficacy and specificity across ESEs mean that often only a well-defined subset of codons can be used to specify the desired ESE.

Unexpectedly, when we derived HPIs for zebrafish amino acids, using a set of ESEs obtained from the same source 24, we found a significant correlation of reverse sign (Dr HPI~rho_x (5'), rho = 0.6, P < 0.003, N = 46; HPI~rho_x (3'), rho = 0.59, P < 0.0033, N = 46). Many experimentally verified ESEs have been characterized as A-rich and C-poor relative to the background frequency of these nucleotides in coding sequence. Whilst we found this to be the case for putative human ESE motifs not shared with zebrafish (A, 47.38% (ESE) versus 25.57% (exonic); C, 15.28% versus 25.99%, N(ESE) = 204), and for ESEs present in both species (A, 50% versus 25.57%; C, 6.37% versus 25.99%, N = 34), unique zebrafish ESEs (that is, ESEs not present in human) from this dataset were unusually enriched in C (39.47% versus 25.99%, N = 288) and relatively poor in A (18.40% versus 25.57%). Although one would expect ESE motifs to vary across taxa, the discrepancies are so pronounced as to sit awkwardly next to the substantial similarities in amino acid trends (Tables 1 and 2). One criterion used by the Burge group 25 to identify candidate ESE motifs was for such motifs to be more common near weak versus strong splice sites. Therefore, one possible explanation is that C-richness is a characteristic of zebrafish ESEs near weak splice sites but not generally, so that the predicted ESEs are not representative of ESEs across the zebrafish genome. Alternatively, comparatively lower quality of the, then recent, zebrafish genome build might be responsible for the divergent results. A re-examination of these putative zebrafish ESEs with an updated genome build may be worthwhile.

To further advance the hypothesis that gradients in amino acid abundance near exon-intron boundaries are a critical feature of exon ends in metazoans, we examined the degree of amino acid conservation as a function of distance from the boundary. For three pairs of species (S. cerevisiae-Saccharomyces castellii, D. melanogaster-Drosophila pseudoobscura (Dps); C. elegans-C. briggsae) sets of orthologous internal exons were derived from various sources and aligned at the amino acid level (see Materials and methods). Mirroring results from a comparison of human-mouse orthologues 2, we found strong and highly significant positive correlations of strikingly linear character (Figure 4) between distance from the boundary and amino acid substitution rate for the Drosophila and Caenorhabditis pairs, whilst proximity to the boundary did not appear to confer a higher level of amino acid conservation in the Saccharomyces comparison. Restricting the analysis to exons of at least 70 codons in length, we obtained qualitatively equivalent results (Drosophilae 5', rho = 0.53, P < 0.002, N = 3,690; Drosophilae 3', rho = 0.77, P = 9.70E-07, N = 3,690; Caenorhabdites 5', rho = 0.74, P = 2.33E-06, N = 6,273; Caenorhabdites 3', rho = 0.58, P = 4.5E-04, N = 6,273). This restriction ensures that all exons contribute an approximately equal share of information to each codon position from the boundary and eliminates the potential confounder that short exons might, for reasons unrelated to splicing, feature more frequently in highly conserved genes and create misleading trends by virtue of their disproportionate contribution to substitution rate information closer to the boundary.

Given that the set of aligned Saccharomyces exons consisted entirely of terminal exons (see Materials and methods), we repeated the analysis for a set of 5,352 orthologous pairs of terminal exons from our Drosophila dataset in order to rule out that differences are caused by any special characteristics of terminal exons. Correlations observed for terminal exons closely resemble those for internal exons (5', rho = 0.83, P = 3.8E-07; 3', rho = 0.75, P = 1.95E-06), alleviating any such concerns.

The above results appear consistent with greater functional significance of boundary-proximal amino acid composition in metazoans, proposed to be at least in part owing to their more extensive utilization of exonic splice regulatory sequences. However, after repeated (k = 10,000) random sampling of 90 aligned terminal exons from the Drosophila dataset and subsequent statistical analysis, we cannot reject the possibility that the Saccharomyces statistics were sampled from the same underlying distribution (Additional data file 9), implying that differences in conservation near exon-intron boundaries cannot be ultimately established from the data at hand.

Having detected higher levels of amino acid conservation near exon-intron boundaries, we expect genes with a high proportion of sequences near boundaries ('flank-heavy') to evolve more slowly. This is indeed what we found when we considered K_A as a function of the proportion of sequence within 70 bp of the boundary (Drosophilae, rho = -0.26, P = 2.2E-16, N = 4,132; Caenorhabdites, rho = -0.08, P = 6.18E-09, N = 5,248; Figure 5). We report K_A rather than K_A/K_S, more commonly used as a measure of selection on protein sequence, because the underlying premise of K_A/K_S, namely that K_S reflects neutral rates of evolution, is violated for sequence encoding ESEs 26.

The results are not qualitatively affected by contracting (50 bp) or expanding (100 bp) the region considered to constitute the boundary flank (Additional data file 10). Focusing on the terminal bins in Figure 5a, it appears that between D. melanogaster and D. pseudoobscura a gene with less than 10% of coding sequence near an exon-intron boundary evolves, on average, almost twice as fast (mean K_A = 0.195) as a gene with more than 70% of boundary-proximal sequence (mean K_A = 0.099). Discrepancies in evolutionary rate between 'flank-heavy' and 'core-heavy' bins appear less marked for the nematode pair (mean K_A (%CDS near boundary >0.9) = 0.12; mean K_A (%CDS near boundary <0.3) = 0.18). However, Figure 5b suggests that this is principally owing to curiously elevated levels of conservation for genes with a small proportion of sequence near the boundary, that is, genes with very large exons, a feature we did not encounter in the analysis of either insect (Dm-Dps) or mammalian (Hs-Mm) orthologues 2.

Importantly, this anomaly highlights a more general reservation, namely that any measure capturing the proportion of sequence near the boundary will strongly covary with exon length, which in turn might covary with underlying functional determinants of evolutionary rate entirely unrelated to splicing control. Thus, in order to control for any putatively distorting effects of functional class on K_A, we employed the following strategy: For each aligned gene, we concatenated the flanking regions of all exons, 5' and 3', defined as the first 72 bp bordering the exon-intron junction of trimmed exons. By implication, genes with no exon larger than 144 bp had to be excluded from this analysis. Concurrently, we concatenated the core sections of all exons of sufficient length in the respective gene, defined as the sequence block enclosed by the two flanking regions. As accurate estimation of K_A probably requires a minimum of 100 codons, we further restricted analysis to those genes with at least 300 bp in the concatenated flanks and in the concatenated cores of exons. For each gene meeting the above criteria we then determined the rates of amino acid evolution in the concatenated core sections (K_Ac) and flanking sections (K_Af). We find that more Drosophila orthologous genes than expected by chance have faster evolving core regions (median (K_Ac - K_Af)/K_Af) = 0.14, Wilcoxon signed rank test P < 0.0001, N = 1,237; Figure 6), consistent with the evidence, presented above, for additional sequence constraint operating on flanking regions. A significant tendency towards more rapid evolution in core sections is also evident when we confine the sample to genes with at least 600 bp in flanking as well as core regions (median (K_Ac - K_Af)/K_Af) = 0.14, Wilcoxon signed rank test P < 0.0001, N = 785). Despite exhibiting the expected shift towards average higher K_A in the core of exons, this trend is much less pronounced than in a previously reported comparison of human-mouse orthologues (median (K_Ac - K_Af)/K_Af) = 0.68, Wilcoxon signed rank test P < 0.0001, N = 360; Figure 6c, and see Parmley et al. 2 for details). Curiously, for the nematode pair, we find significant evidence for a reverse correlation (300 bp, median (K_Ac - K_Af)/K_Af) = -0.07, Wilcoxon signed rank test P < 0.0001, N = 1,102; 600 bp, median (K_Ac - K_Af)/K_Af) = -0.014, P < 0.038, N = 496), that is, in the majority of genes, flanking regions evolve at a marginally higher rate than core regions.

Parmley et al. 2 recently presented evidence that, in mammals, amino acid usage in the vicinity of exon-intron boundaries is affected by factors unrelated to protein function but to sequence-based information required for correct splicing. The objective of the present study was to elucidate whether such requirements have left an evolutionary imprint on exonic sequence composition across a phylogenetically diverse set of species. To this end, we systematically compared trends in relative amino acid abundance near exon-intron boundaries in 12 eukaryotic species. Our analysis revealed that preference for or avoidance of certain amino acids near boundaries is a common phenomenon among metazoan species but is not unique to metazoans. More amino acids show skewed usage in species where a greater problem identifying intron-exon boundaries is to be expected, that is, those with large and numerous introns. Notably, this includes the basidiomycete C. neoformans, suggesting that exonic splicing regulation might be a generic characteristic of species with complex pre-mRNA structures rather than absent from the fungal kingdom by virtue of phylogeny. Preference patterns show unmistakable signs of conservation along several dimensions: composition, relative strength, and directionality. The concordance in directionality (whether an amino acid is preferred or avoided) is particularly impressive in that we observe many commonalities with the mammalian pattern even in only distantly related species.

We do not claim that the systematic patterns we observe are solely caused by a selected preference for codons involved in ESEs. In fact, composite trends are almost certain to be the result of multiple functional constraints, including the need to avoid intron-specific enhancer motifs (for example GGG in mammals 25) as well as motifs that would disrupt exon recognition. Furthermore, abundance trends could partially be the result of cryptic splice site avoidance as suggested by Eskesen and colleagues 27. However, many of the trends observed - for example, cytosine avoidance near boundaries - are not predicted by this model 211.

Introns associate non-randomly with the codon in direct proximity to the splice site in a phase-specific manner, an observation often described as insertional preference 28. Trimming and elimination of the first full codon should guard against picking up such insertional preferences or an extended splice site consensus. We cannot rule out that some boundary-proximal codons have slipped into our dataset owing to poor splice site annotation. However, it must be pointed out that this reservation applies only to the subset of amino acid trends that show biased usage directly adjacent to introns and might be more relevant to the interpretation of local discontinuities (Additional data file 11). Also, if the above-mentioned explanations were of major relevance, we would expect cryptic splice site avoidance, insertional preference, and (to a lesser extent) poor splice site annotation to cause similar patterns in ascomycetous yeasts, in particular S. pombe, for which a dataset of reasonable size is available. This is not the case.

Establishing to what extent these trends are caused by preference for ESEs will ultimately depend on characterizing species-specific catalogues of ESE/Exonic splicing silencer (ESS) motifs together with their corresponding trans-factors and relating these to the observed spectra of preferred/avoided amino acids. This work, in particular relating to tissue- and stage-specific splicing patterns, is still in its infancy 29, the catalogues currently available restricted to a small number of vertebrates and yet to be fully verified experimentally 3031.

However, the dearth of significant trends in S. cerevisiae and S. pombe strengthens the proposition that preference trends principally reflect requirements to accommodate exonic splicing regulators. Although the S. cerevisiae genome codes for an SR protein kinase (Sky1p) with the capacity to phosphorylate mammalian arginine-serine rich (RS) domains, the likely endogenous substrate (the SR protein-like Npl3p) does not appear to be involved in pre-mRNA splicing 332. Importantly, no splicing factors homologous to metazoan SR proteins have been discovered in S. cerevisiae 14, consistent with the classical view that splicing in budding yeast is regulated intronically. This is further consistent with the observation that splice site consensus is generally highly conserved, especially 5', much more so than in other species, including C. neoformans (Additional data file 5). The fact that our analysis revealed a significant 3' trend for the two-fold block of leucine (L2) might hint at the presence of recognition motifs in yeast exonic sequence. However, at present there is no evidence supporting the regular involvement of an ESE-like binding motif in S. cerevisiae splicing and alternative explanations should be considered.

Splicing in S. cerevisiae is moderately common in quantitative terms because many highly expressed genes, notably encoding ribosomal proteins, contain introns, so that over 25% of the mRNA population are spliced 33. However, in over 6,000 S. cerevisiae genes we find less than 300 introns in total, so that splicing can hardly be considered a processing stage representative on a genome-wide scale. In contrast, splicing is much more prevalent in S. pombe where approximately 40% of genes contain introns 34. Basal splicing proteins show an enhanced similarity to their mammalian homologues and two SR protein homologues (Srp1p, Srp2p) have been identified 353637. Unlike in budding yeast, there is recent evidence that Srp2p binds to specific exonic elements and interacts with the fission yeast orthologue of human splice factor U2AF 38. Why then, given that SR protein-ESE-like interactions seem to exist in S. pombe, do we not find any trends for amino acid or codon preference in this species? We suggest that trends may be lacking for two reasons. Firstly, given the comparatively low level of splice site consensus degeneracy, a minimal number of ESEs might be sufficient to ensure correct splicing. On a genomic level, we might then fail to register biased abundance patterns on the spatial scale investigated in this study. Secondly, for clear-cut preference trends to evolve, a minimum level of splice-regulatory complexity might be required. This fits with our observation that more amino acid trends are observed in species with complex, intron-rich gene structures, including the yeast C. neoformans (Additional data file 6). Further, alternative splicing contexts, where regulatory elements frequently compete for precedence if arranged close to each other, could be envisaged as an evolutionary pressure initially driving the diversification of ESEs and corresponding trans-factors, thereby creating an environment in which strong trends might be required to attract or repel the correct set of trans-factors, both for constitutively and alternatively spliced genes. Consistent with this hypothesis, reports of alternative splicing in S. cerevisiae 39 and S. pombe 40 are restricted to singular cases, for which functionality of the recovered alternative splice products remains to be shown 41. However, attempts to link diversity and density of ESEs to alternative splicing have so far yielded ambiguous results 42.

The absence of preference patterns in ascomycetous yeasts has an important practical implication. Finding amino acid trends to be abundant near exon-intron boundaries can be regarded as evidence for exon-based splicing regulation, without prior knowledge of specific binding motifs or trans-factors, although failure to detect such trends is insufficient to rule out interaction between exons and auxiliary proteins in the splicing process (compare S. pombe).

The fundamental deviation from the 'mammalian pattern of directionality' shown by the 5' amino acid trends in nematode exons (Table 1) might, at first sight, be unexpected. There are extensive homologies between vertebrate and nematode basal splicing machineries on the protein level 13. Furthermore, splicing in SR protein-depleted cells of the Caenorrhabditis relative Ascaris lumbricoides can be rescued by adding SR proteins derived from non-nematode (HeLa) whole cell extracts, supporting at least a minimum degree of functional overlap 43. Thirdly, the high level of conservation between SR and SR-like proteins identified in each species explicitly includes the RNA recognition motifs, tentatively suggesting similar binding specificities 44.

There is, however, one feature of the nematode splicing process that sets it apart from the other species in our sample: a substantial proportion (approximately 70%) of C. elegans (and C. briggsae) genes are trans-spliced 45. In this process a short (22 nucleotide) 5' small nuclear RNA (snRNA) fragment, the spliced leader, which is transcribed from a different genomic locale, is added at the 5' end of the pre-mRNA 46. It would, we suggest, be highly disadvantageous for this trans-splicing machinery to act at the 5' end of exons where cis-splicing should occur. Indeed, were trans-splicing to occur where intron removal should take place, a gene would, in effect, be broken in two. Thus, we suggest that 5' ends of internal exons have evolved to ensure that they do not attract the trans-splicing machinery. Given that this machinery is ubiquitous in a cell, all 5' ends of internal exons, be they from trans-spliced genes or not, should be equally under pressure to avoid trans-splicing where cis-splicing should happen. Consistent with this expectation, the trends seen at 5' and 3' ends in internal exons are the same in genes from operons and those not in operons (data not shown). Interestingly, information content at 3' splice sites in nematodes is strikingly higher than in other species (Additional data file 5), as previously observed 47, further supporting the idea that splicing regulation in nematodes is unusual in its asymmetry.

What might be the proteins involved in trans-splicing? There is good evidence that several stages of the trans-splicing process are, like cis-splicing, critically supported by SR proteins 4348. Furthermore, whilst mammalian and Ascaris SR protein extracts are equally efficient in catalyzing cis-splicing in vitro, Ascaris SR protein extracts engender an approximately five-fold higher trans-splicing activity 43. Although the use of whole cell extracts in these experiments precludes an analysis of the differential contribution of individual SR proteins, these observations are consistent with the hypothesis that a subset of splice-regulatory proteins in these species is dedicated to trans-splicing.

Given the above, we envisage trans-splicing specific SR and other proteins to interact primarily with intergenic sequence upstream of the first exon of the pre-mRNA to provide further guidance for the trans-splicing apparatus or mediate other functions crucial to trans-splicing, such as protecting downstream RNA from degradation 4549. A prediction derived from this model is that we should find in nematodes proteins participating in trans-splicing that bind to nucleotide motifs depleted of codons from amino acids avoided near the 5' end of exons.

Owing to their deviant 5' trends, nematodes stand out in another aspect of systematic amino acid biases. Parmley et al. 2 observed no significant differences in preference trends between 5' and 3' ends of exons in mammals. Similarly, approximate symmetry has been reported for ESE distribution in human exons 30. Conversely, standardized major axis regressions 2122 strongly suggest that nematodes do not conform to a symmetric pattern of preference trends.

An assessment of this situation very much depends on how we expect ESE-guided splicing regulation to work on a mechanistic level. If SR proteins are assumed to interact directly with specific components of the basal splicing machinery, as is probably the case for U2AF 3, we would not automatically expect the same ESEs (and by implication amino acid trends) to be represented at similar frequencies 5' and 3' where different spliceosomal proteins are present. Predictions of whether symmetry might be of functional relevance, however, especially for scenarios of indirect interaction, cannot be derived from the data at hand.

Confidence intervals in our exploration of symmetry are large so that we cannot ascertain that symmetry is a dominant pattern throughout our species sample. However, some best estimates of SMA slopes (β) are tantalizingly close to perfect symmetry (Mm, β = 0.9907; Hs, β = 1.0362; Dr, β = 1.0439; Ag, β = 1.0788; Table 5), warranting more detailed examination of this potentially functional signature in the future.

Consistent with the proposition that trends in relative amino acid abundance are functionally important, we observe lower rates of nonsynonymous evolution near exon-intron boundaries in insects (Dm-Dps), nematodes (Ce-Cb) and mammals (Hs-Mm), indicative of higher selective constraint in this region. Furthermore, the proportion of coding sequence that is located near boundaries is a partial predictor of K_A (Figure 5). Genes with a higher share of sequence partaking in exon flanks tend to show reduced rates of evolution. Nematode genes, again, stand out in that they do not conform to the negative linear relationship between K_A and flank-heaviness found in other species pairs (Hs-Mm and Dm-Dps), but show unexpectedly high levels of conservation for genes with very large exons. The causes for this currently remain elusive. Similarly, we would not have predicted that in worms gene-specific differences between evolutionary rate in the flanking and core sections of exons are biased (if only slightly) towards more rapid evolution of flanking regions. However, the distribution of core-flank evolutionary rate differentials in worms appears comparable to the one for flies, a higher median evolutionary rate of core regions in the latter notwithstanding (Figure 6). Human-mouse orthologous genes on the other hand show a much more dramatic distributional shift towards faster evolution in exon cores (see distributions in Figure 6). Between-taxa differences in gene composition, especially relating to the presence of more and longer introns in mammals, might account for these differences: on a speculative note, information necessary to distinguish an exon from surrounding non-coding sequence might require a unique degree of conservation under these circumstances, perhaps severely restricting the leeway for nonsynonymous changes to occur in flanking regions. Alternatively, restrictions imposed by our experimental set-up, especially relating to minimum sequence length requirements, might have resulted in the selection of gene sets with divergent splicing characteristics in the different species pairs. We leave a closer dissection of these questions to further analysis.

For 12 species (human, mouse, zebrafish, C. elegans, C. briggsae, A. gambiae, D. melanogaster, A. mellifera, A. thaliana, S. cerevisiae, S. pombe, C. neoformans) we established individual exon datasets derived from a small number of databases (Additional data file 12). Pre-established CDS tracks were followed in all but three cases (At, Sp, Cn), for which annotated chromosome/scaffold sequences were downloaded from the relevant database and exons extracted subsequently. Exons with identical locus IDs were then sorted into individual files, only retaining files with at least one internal exon. All locus files were subsequently checked to ensure coding sequence started with ATG, finished with a stop codon (TAA, TAG, TGA), had no internal stop codons, and was a multiple of three nucleotides. Locus files where one of the above prerequisites was violated were removed from the final dataset. We also eliminated exons containing one or more ambiguous nucleotides ('n'). The remaining exons were trimmed so that the first nucleotide was the first nucleotide of the first complete codon and the last nucleotide the last of the final complete codon. Then, we discarded all terminal exons to obtain the final exon sets. Gene models from which exons were derived are provided in Additional data file 13.

After splitting individual exons in half to ensure that no codon featured in both 5' and 3' analyses, we considered the trend in usage of each amino acid as a function of the distance from the boundary up to a maximum distance of 34 codons. Importantly, the codon in direct proximity to the boundary was also eliminated.

We then calculated Spearman rank correlations (rho) between the distance from the boundary (5' or 3') and proportional usage of the amino acid (that is, in proportion to the number of residues at that given distance) for the remaining 33 data points for each species. The three six-fold degenerate amino acids we split into blocks of four and two (that is, 'S4' signifies, TCA, TCC, TCG and TCT, while 'S2' signifies AGC and AGT). In relevant circumstances, the two-fold and four-fold blocks were treated as separate amino acids, yielding a total of 23 amino acid groupings.

For each amino acid grouping independently we fitted unweighted linear models and extracted the slope of the regression line to be used as a basic measure of the strength of individual preference trends. Note that a negative rho/slope implies an amino acid that is preferred near boundaries and a positive rho/slope implies a tendency to be avoided. Unless otherwise stated, results are reported as significant only if they remain significant after correction for multiple testing (see Results for adjusted P-values).

For the most part, trends are approximately monotonic and linear and hence adequately captured by simple linear models. For certain amino acids, departures from linearity, some recurrent across species and typically highly localized, do exist however. Unusual U-shaped 5' trends for proline, originally noted for human and mouse by Parmley et al. 2, are also present in other species (Ce, Dr). Further, some amino acids, notably isoleucine and the two-fold block of leucine, are disproportionately preferred in direct proximity to the boundary (after trimming) at 3' exon ends in several species. 'Popping out' from otherwise linear trends (Additional data file 14), these patterns are perhaps caused by participation of the relevant codons in an extended splice site consensus relevant for U5 snRNA-mediated exon joining (see Additional data file 11 for a more detailed discussion of recurrent, locally confined preference/avoidance patterns and potential functional explanations). As a corollary of discontinuities more generally, comparative interpretation of slope coefficients as an index of relative strength ought to be done with care. In particular, our rank ordering of slopes derives its value from providing another dimension through which congruence in preference spectra can be asserted, rather than being easily translated into differential functional impact on a mechanistic level.

Given the small number of internal exons in S. cerevisiae (only eight genes have more than one intron), we decided to include terminal exons in the final dataset (417 exons) for this species. The one end of each terminal exon that did not border the intron was excluded. Otherwise, the removal of irregularities (internal stop codons and so on) proceeded as described above. Restricted sample size also indirectly prompted a re-examination of the results obtained from Spearman's rank correlations because the presence of multiple tied ranks led to concerns about the adequacy of this statistic. However, using the more appropriate Kendall's tau statistic did not return any qualitatively different results.

For 5' and 3' datasets independently, Spearman's correlations were computed between the previously derived slope coefficients of all 23 amino acid groupings for every possible metazoan species pair. Ninety tests (with the number of species N = 10, N^2-N = 90) were carried out and significance threshold adjusted accordingly (P = 0.05/90 = 5.56E-04). We initially included both yeast species in the analysis but, as expected from the absence of significant individual amino acid trends, we found no significant correlations for the global amino acid set (data not shown). No loss of relevant information is incurred whilst clarity of presentation is enhanced when these species are excluded from the analysis and, in particular, the accompanying table (Table 5).

Covering all 23 amino acid groupings Spearman's rank correlations were computed between 5' and 3' trends within each species (N = 12, P = 0.05/12 = 4.17E-03).

SMA regressions were computed in R using the SMATR package 2122 applying standard confidence limits (95% CI). As symmetry of the type x = y was to be tested, the regression line was forced through the origin. SMA regression requires estimates of the slope of the regression line to have a consistently positive or negative sign so that the major and minor axes can be identified unambiguously. This is not the case for either A. thaliana or C. neoformans, which are hence not amenable to this type of analysis and were not included. Further, residual distribution for S. cerevisiae shows significant deviation from normality so that results for this species should be interpreted with care.