An important approach to identifying functionally important AS is to focus on isoforms common to orthologous genes in two species (11–14). This method does not eliminate the possibility of false positives representing aberrant events that have occurred in both species, or false negatives resulting from insufficient data or lineage-specific functional AS (15), but it is probably the most reliable current computational approach to identify functional AS isoforms on a large scale. Comparisons of conserved and nonconserved AS reveal systematic differences, with nonconserved AS exons appearing more similar to constitutively spliced (CS) exons in several respects (16); this was interpreted as indicating that the nonconserved cases are likely enriched for aberrant events (16). Conserved AS exons display increased exonic and flanking intronic sequence conservation compared to CS exons (12–14), indicating that they have more bases under purifying selection, in accord with experimental work finding a high density of splicing enhancers and suppressors associated with AS regulation, in both exons and introns (refs. 17 and 18 and references therein). It is also known that, in comparison to CS exons, AS exons tend to be smaller and have weaker splice site signals (14, 17, 19–23).

Here, we attempt a more detailed characterization of evolutionarily conserved AS. We focus on exon skipping (the most frequent type of AS in mammals; ref. 13) and classify AS cases by whether they are predicted to result in isoforms subject to NMD and whether the skipped “cassette” consists of single or multiple exons. Our data set is considerably larger than in earlier studies, which permits more powerful modes of analysis and several unique insights. Two methodological innovations in our analyses include a nonparametric likelihood-based approach to identify conserved AS cases that likely represent aberrant splicing that has occurred by chance in both species (if unremoved, such cases can bias analysis results), and a simple quantitative measure, the “effective number of conserved nucleotides” (N_c) to estimate the overall amount of evolutionarily conserved regulatory information associated with particular exons.

We find that a significant number of apparent NMD-inducing AS isoforms are conserved in mouse and human and display increased sequence conservation relative to CS exons, supporting Lewis et al.'s suggestion (6) that some NMD-inducing AS is functionally important (although the proportion of overall cases we find that are NMD-inducing is smaller than in ref. 6). We show that the increased sequence conservation that had previously been observed in AS exons mainly affects synonymous sites, such that AS exons have substantially lower synonymous substitution rates and slightly lower nonsynonymous substitution rates than CS exons, likely reflecting the presence of exon splicing enhancers and suppressors. (Iida and Akashi had previously noted a lower synonymous substitution rate in AS exons, but proposed that it was due to selection on codon usage; ref. 24). Small CS exons also exhibit increased conservation in synonymous sites (although to a lesser degree than AS exons), suggesting that a higher density of signals is required to facilitate their splicing as well.

Identification of Conserved AS and CS Isoforms. For details of the procedures and a schematic overview, see Supporting Text and Fig. 7. We aligned 38,258 human near-full-length protein-coding cDNAs to the human genome sequence; after filtering out unspliced or problematic alignments, there were 28,172 aligned cDNAs (called “reference cDNAs” below) in 14,394 distinct genes. We obtained 6.26 million alignments of human ESTs and cDNAs to the human genome from the University of California, Santa Cruz, Genome Bioinformatics Site (25) (July 23, 2004, release), which reduced to 1.69 million alignments after filtering. The same procedures were used to obtain 26,239 mouse reference cDNAs for 12,499 distinct genes and 5.23 million alignments (0.95 million after filtering) of mouse ESTs and cDNAs (July 23, 2004, release) to the mouse genome.

For each EST or cDNA that was not itself full length as judged by comparison to the reference cDNA, we inferred a putative reconstructed full-length structure by combining with the upstream and downstream structure from the corresponding reference cDNA. The reconstructed coding sequence was then conceptually translated to identify reading frame and the location of the stop codon. If the latter was ≥50 bp upstream of the final exon–exon junction, the corresponding transcript was designated NMD-inducing (26–28). (It should be noted that this 50-bp rule is imprecise; although it accords with most available experimental data, it may sometimes be violated and the efficiency of NMD is in any case somewhat variable; ref. 29.)

Instances of exon-skipping AS were identified by comparing the genomic coordinates of pairs of reconstructed cDNAs from the same gene. We used the University of California, Santa Cruz, blastz (30) alignments of the mouse and human genomes to identify orthologous exons having identical predicted status (CS, or frame-preserving, non-frame-preserving non-NMD, or NMD-inducing AS) in both species. The data set of mouse–human conserved AS isoforms and their characteristics is given as Table 4.

Splice Site Scores. From the aligned human reference cDNAs, a nonredundant set of internal exons was identified and grouped into four bins based on the G+C content of the flanking 200-kb region. Gene-specific log odds (base 2) weight matrices (31) for a window of 15 intronic and five exonic bases around the splice site were computed for donor and acceptor sites, by using bin and position-specific “foreground” frequencies of nucleotides from the sites in the bin, and gene-specific background frequencies determined from the transcribed region of the gene. A similar procedure was used to find weight matrices for mouse splice sites.

Sequence Conservation. We used blastz-aligned mouse and human orthologous regions (exons, splice site 20-bp windows as defined above, or intronic 20-bp windows immediately adjacent to the splice site windows) to compute percent identity (PI) with respect to the human sequence. Unaligned positions were considered to be mismatches, and gaps in the mouse sequence were disregarded. Synonymous and nonsynonymous substitution rates (Ks and Ka) and their confidence intervals were estimated by the Pamilo–Bianchi–Li algorithm (32, 33) as implemented in the mega2 package (34). The conserved (non-splice site) intronic region was determined by assigning columns in the blastz alignment a score of +1 if the bases were identical, and –3 otherwise, and finding the highest scoring segment that starts at the end of the splice site window (i.e., at base 16 from the site) and extends further into the intron (truncating the search at the start of the next exon).

Effective Number of Conserved Nucleotides. The conserved intronic region identified as above likely consists of a mixture of neutrally evolving bases and functional (enhancer or silencer) sites, so its length does not provide an accurate measure of the amount of conserved information. We propose the following simple method to quantitate this information, given a pair of orthologous aligned sequences. A region of size n_o aligned bases with average (fractional) identity r_o between the two sequences is assumed to consist of n_n neutrally evolving positions with average identity r_n, and n_c positions under purifying selection with average identity r_c > r_n. Then, n_cr_c + n_nr_n = n_or_o. Because n_n = n_o – n_c we have n_c = n_o(r_o – r_n)/(r_c – r_n). We estimate the neutral value r_n from intergenic and intronic regions within a 200-kb window centered on the exon, separately for each gene (to allow for variation in the rate of neutral evolution across the genome; ref. 35). Although r_c is unknown (and may vary for different functional motifs), note that the effect of using a different fixed value for r_c is simply to multiply n_c by a constant factor. Therefore, we can for simplicity take r_c = 1.0. The value of n_c obtained for this r_c is the equivalent number of perfectly conserved nucleotides that would yield the same overall conservation level. We designate this N_c and call it the “effective number of conserved nucleotides.” A similar estimate for exonic synonymous sites is obtained by using a more complex procedure (see Supporting Text).

Exon Exclusion Rate. Exon exclusion rate was computed as the number of ESTs or cDNAs supporting the exclusion isoform, divided by the total number of ESTs or cDNAs spanning a genomic region that contains the relevant exons; this constitutes a rough average over tissues according to their representation in the EST database, and in particular does not distinguish intertissue from intratissue differences in isoform frequencies.

Table 1 confirms that frame-preserving AS is the most prevalent type, accounting for 62% (920 of 1,478) of conserved AS cases. However, it is striking that 38% of conserved cases do not preserve frame, and 24% (361 of 1,478) appear to be subject to NMD. Use of a filter to remove likely aberrant splicing cases (see below) reduced this value to 21% (192 of 900 single-exon skipping cases). The latter figure is somewhat lower than the estimate of one-third in ref. 6, derived from human EST data without regard to conservation; however, it should be noted that our results suggesting underrepresentation of NMD-inducing AS isoforms (see below) imply that prevalence of NMD-inducing AS may be higher than we currently estimate. An earlier study of conserved AS based on substantially smaller data sets obtained a higher estimate (73%) for the fraction of frame-preserving AS (16); this probably reflects the fact that underrepresented NMD-inducing isoforms are more likely to be revealed as we obtain increasing amounts of data.

Roughly 14% (204 of 1,478) of AS cases involve multiple exons. One-third (29 of 88) of the frame-preserving multiple-exon skipping cases involve a combination of exons that individually have lengths not a multiple of 3, and which would therefore be frame-shifting if individually deleted.

Sequence Conservation by AS Category. Previous studies (12–14, 24) have shown that evolutionarily conserved AS is accompanied by increased sequence conservation in both exons and introns compared to CS exons, likely reflecting the presence of splicing regulatory elements subject to purifying selection. We verified that each of the major AS classes in Table 1 exhibits significantly elevated sequence conservation in exons and 20-bp windows surrounding splice sites, relative to CS exons (Figs. (Figs.11 and 9). Using a measure of the total amount of conserved sequence information (the “effective number of conserved nucleotides,” N_c), we verified that the major AS classes also have significantly greater N_c than CS exons (Figs. (Figs.11 and 9). This finding strongly suggests that, in each of the AS classes, the splicing is regulated, and hence that the variant forms are, in general, functionally important rather than aberrant events. In particular, conserved NMD-inducing AS isoforms may help regulate gene expression by reducing protein levels, as suggested by Lewis et al. (6). NMD-inducing inclusion isoforms represent an interesting subclass comprising about one-quarter of the NMD cases, with the role of the included exon apparently being to mediate destruction of the message which contains it, thus acting as a “switch” to down-regulate the gene.

Discrimination of Misclassified Exons. As has been widely recognized (9, 10, 36), EST and cDNA databases contain numerous variants that represent aberrant splicing rather than functionally important AS. Our evolutionarily conserved AS cases could include two types of such variants: aberrantly included cryptic exons that derive from neutrally evolving intronic DNA, and exons that are functionally CS but have been aberrantly excluded from some transcripts. However, there is an asymmetry in the expected frequencies of these events. Cryptic exons should be rare in our data set given the requirement for orthologous isoforms in both species, because the high neutral sequence divergence between mouse and human (>30%) makes cryptic inclusion events in precisely orthologous locations in both species unlikely. In contrast, depending on rates of aberrant skipping and EST sampling depth, occasional aberrant skipping of the same functionally CS exon might well occur in both species. In fact, a consequence of the “exon definition” model (23) is that aberrant skipping is likely to be a relatively common type of splicing error, because failure at any step in exon recognition would tend to cause the exon to be skipped. With enough data, aberrant splicing events may cause virtually all genes to appear alternatively spliced (5), so conserved AS cases that represent congruent skipping errors in the two species will become increasingly frequent as the EST database expands.

Consistent with this possibility, scatterplots (Fig. 10) indicate that some AS exons with low exclusion rate appear more similar to CS exons than to other AS exons with respect to sequence conservation, splice site score, and/or size, suggesting that the few exclusion transcripts for these exons may represent aberrant events. Moreover, some exons classified as CS in our data set may represent AS exons for which the exclusion isoform does not yet have EST support in either species.

Consequently, as described in Supporting Text, we developed a log-likelihood-based discriminator LL(E) based on seven characteristics (exon size, combined splice site score, whether the exon size is a multiple of three, and four measures of exonic and intronic sequence conservation) to help distinguish “misclassified” exons. (We regard an exon as correctly classified as AS only if both variants are functionally important in the sense that selection acts to preserve them, a criterion that admittedly may not be possible to reliably determine without substantial experimental work, but for which a reasonable circumstantial case can often be made based on sequence conservation and other features.) AS exons should generally have positive values for LL(E), whereas CS exons should have negative values. Roughly 29% of the 1,274 single-exon skipping AS exons in our data set have LL(E) values <0, reflecting sequence characteristics indicative of CS exons, whereas 11% of the 14,368 CS exons have LL(E) values ≥0, reflecting sequence characteristics indicative of AS exons (Fig. 11). Fig. 11 shows evidence of a separate “peak” of AS exons with LL(E) scores <0; and these AS exons have, on average, a lower exclusion rate than AS exons with positive LL(E) scores [0.169 for AS exons with LL(E) < 0 vs. 0.453 for AS exons with LL(E) ≥ 0], consistent with the possibility that their exclusion transcripts represent aberrant splicing, although some may represent true AS cases at the distribution tail. Given that the LL(E) score distributions of true AS and CS exons probably overlap somewhat, the discriminator is not perfect. Many of the analyses we perform below could be biased either by including misclassified cases, or by excluding true cases whose values are at the distribution tail of the characteristic being analyzed. To minimize both biases to the extent possible, for each analysis we developed an LL(E) score based on an appropriate subset of the seven characteristics (i.e., not including parameters related to the ones being analyzed), and used this to filter out likely misclassified cases. The filters, and which analyses they were used in, are shown in Table 5. Analogous classifiers (however, based on an assumption that all conserved AS cases are genuine, and using different statistical principles) have recently been developed by several other authors (14, 37–39); their characteristics are compared in Table 6.

Exonic Sequence Conservation Patterns. To illuminate patterns of exonic sequence conservation, we compared Ks (synonymous substitutions per synonymous site) and Ka (nonsynonymous substitutions per nonsynonymous site) in filtered CS and frame-preserving AS exons. Increased sequence conservation in AS exons derives primarily from synonymous sites: average Ks in AS exons is one third that in CS exons (0.166 vs. 0.510, P < 10^–10), whereas Ka is only moderately lower (0.0648 vs. 0.0808, P < 10^–10). Although the conservation is slightly stronger near splice sites, it is dispersed broadly over the entire exonic region (Fig. 2A). These observations support the idea that a substantial fraction of the exonic sequence in AS exons lies within exon splicing enhancer and suppressor signals (17, 18, 40–42), and that, in a sequence whose primary function is encoding a protein, such signals are more easily accommodated by placing critical motif sites at synonymous positions.

In general, small exons are thought to be poorly recognized by the splicing machinery, possibly as a result of steric hindrance between snRNPs and other splicing factors (22, 43); presumably to compensate for this, small CS exons tend to be accompanied by stronger splice site signals and splicing enhancer motifs (22, 44). Because AS exons tend to be small (21, 23), it is important to consider the possibility that the increased sequence conservation in AS exons is due solely to their size. Consequently, we examined Ks and Ka for CS and AS exons by exon size (Fig. 2B). Small (≤50 bp) CS exons indeed have 18% lower Ks (0.415 vs. 0.506, P = 4.8 × 10^–10) than medium (100–150 bp) CS exons. However, AS exons have substantially lower Ks than CS exons regardless of size (Fig. 2B), suggesting that they carry additional information beyond that required to compensate for the difficulty in splicing small exons. On average, short exons have lower synonymous N_c than larger exons (Fig. 3), suggesting that their increased conservation reflects signals being packed more tightly into a smaller space (rather than there being more of them in absolute terms).

These findings suggest that synonymous sites are important carriers of splicing regulation information in AS and small CS exons. One implication is that the common practice of using Ks to represent the neutral rate “baseline” for the purpose of assessing selection on protein coding sequences with the Ka/Ks ratio can be misleading for such exons.

Factors Influencing Exclusion Rate. As a preliminary step in understanding the factors controlling the relative frequencies of AS isoforms, we computed correlations of various sequence characteristics with exclusion rate (Table 7), using frame-preserving AS cases because exclusion rates for NMD-inducing AS cases are likely distorted as a result of underrepresentation of the NMD-inducing isoform (see below). After filtering, splice site PI, intronic N_c, intronic PI, strength of splice site signal, and exon size all show significant correlation with exclusion rate, but exonic PI does not. Figs. Figs.44 and and55 confirm that most of the significant correlations reflect reasonably consistent trends across the range of exclusion rates: increasing exclusion rate is accompanied by decreasing exon size, decreasing splice site signal strength, and increasing numbers of conserved bases (N_c) in the flanking introns. In contrast, splice site and intronic sequence conservation, though significantly stronger than for CS exons, plateau at modest exclusion rates (Fig. 5).

To help interpret these trends we examined the correlation of exon size with other characteristics in CS exons (Table 8). In general the correlations, although highly statistically significant, are relatively weak; however, Fig. 6 demonstrates a stronger splice site signal and elevated sequence conservation in the splice sites in small as compared to medium CS exons. No or marginally significant difference in intronic N_c was observed (Table 8 and Fig. 3).

We can now begin to interpret the exclusion rate trends. A fundamental question is the following: does selection act to produce particular exclusion rates, in which case the trends reflect the mechanisms by which the organism achieves the desired rates (e.g., weaker splice sites and smaller exons to produce a high exclusion rate); or is the observed exclusion rate instead simply a passive result of splice site score and exon size? The answer may depend on the gene to some degree, but overall our results on sequence level selection appear to favor the first possibility: if exclusion rate were not under selection, it is hard to understand why at higher exclusion rates there should be both low splice site scores and strong selection on those splice sites to maintain their sequences (Figs. (Figs.44 and and5).5). Similarly, because smaller size is compensated by stronger splice site signals in CS exons (Fig. 6), in the absence of selection on exclusion rate, one would expect any selective pressure acting on splice site signals for small AS exons to be in the direction of maintaining stronger rather than weaker splicing signals, the reverse of what is observed. So, it appears that it is exclusion rate that is being selected for, and that smaller exon size is one mechanism for achieving a higher exclusion rate. This finding is a bit surprising because one might have expected the size of the included AS exon to be dictated primarily by protein structural requirements, and it raises the interesting possibility that, even in frame-preserving AS cases, the role of the included exon may sometimes be a negative one (disrupting the protein function), analogous to the NMD case (although this cannot always be so, because there is an overall bias in frame-preserving AS toward preserving domain structure; refs. 14 and 45).

We carried out a multivariate regression to derive a linear model to predict exclusion rate as a function of exon size, intronic N_c, splice site score, and intronic PI (Table 2). The model is oversimplified in that it does not take into account the precise mix of enhancer and silencer motifs, and their tissue and/or developmental stage specificity, all of which must be important in determining exclusion rate for individual genes, but it may give some insight into the overall processes involved. The regression results confirm that each of the four factors contributes significantly to the determination of exclusion rate (Table 2), and together account for about one-quarter of its variance (R² = 0.26).

The exclusion rate trends suggest that the organism achieves higher exclusion rates by dispersing the sequence information required for splicing, moving it out of the splice sites and into the introns and exons. Both the splice site signal and exon size play in some sense a “suppressor” role, which is presumably compensated to some degree by exonic and intronic enhancers. It will be of interest to assess the relative frequencies of enhancer and suppressor elements as a function of exclusion rate.

Underrepresentation of NMD-Inducing AS Isoforms. Because they are removed by NMD, NMD-inducing AS isoforms are expected to be underrepresented in the EST and cDNA databases. To test this, we applied our regression model to predict the exclusion rate for each NMD AS case. For the “exclusion isoform NMD” category, the average predicted exclusion rate is 0.239, significantly higher (P < 10^–10) than the observed rate of 0.107. For the “inclusion isoform NMD” category, the predicted rate (0.311) is significantly lower (P < 10^–10) than the observed rate (0.696). In both cases, the direction of the deviation is consistent with the NMD-inducing isoform being underrepresented relative to what is predicted by the regression model, suggesting that this isoform is in fact being removed from the mRNA pool. These trends are reflected in the observed relative frequencies of the different AS groups along the exclusion rate axis (Fig. 12), with inclusion isoform NMD cases being relatively more frequent at high apparent exclusion rates, and exclusion isoform NMD cases more frequent at low exclusion rates. For non-frame-preserving non-NMD cases, the predicted rate (0.263) is slightly less than the observed rate (0.337) (P = 5.2 × 10^–4), suggesting that some inclusion transcripts may be relatively underrepresented. It is possible that these represent instances where NMD occurs even though the 50-bp premature termination codon rule is not met (46, 47), or that they undergo decay via pathways other than NMD, such as a mRNA decay pathway that destroys transcripts with a termination codon in the last exon (48).