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.