Vertebrate genes consist of relatively short exons separated by considerably larger introns. The introns of lower eukaryotes, invertebrates and plants are much shorter. This difference may be explained by the preference for two possible mechanisms for recognition of the exon-intron boundaries by the splicing machinery. In the case of long introns, the exon definition mechanism initially recognizes pairs of splicing sites corresponding to one exon. Vice versa, short introns are recognized by the intron definition that pairs splicing sites across introns 1. Historically, the intron definition mechanism seems to be the ancestral one, whereas exon definition likely is a relatively recent innovation that, in particular, created the possibility of regulated alternative splicing 2.

These models yield different consequences of mutations that destroy splicing sites. Errors in exon definition should lead to exon skipping or, if there are strong cryptic sites, the use of the latter, whereas errors in intron definition should cause intron retention. Indeed, exactly this behavior was observed in vivo and in vitro experiments (reviewed by 1), and in early analyses of disease-causing mutations of human genes 34. These predictions also agree to the distribution of alternative splicing types in different organisms. In vertebrates, where long introns are frequent, the prevalent type of alternative splicing is exon skipping 56, while in plants, where the majority of introns are short, the most frequent type is intron retention 57.

Intron retention is the least studied type of alternative and aberrant splicing. In contrast with other types of alternative splicing, which involve the choice between different splice sites, intron retention represents complete absence of splicing. Some specific features of retained introns have become clear in recent studies of human 89 and plant transcriptomes 10. Retained introns were found to differ from other introns in GC content, that was lower than in exons but higher than in constitutively spliced out introns. Retained introns were shown to be shorter on the average than constitutively spliced out ones and exhibited a tendency to occur in 5'- and 3'-untranslated regions 8910; they also have weaker sites 9.

In several cases intron retention clearly has a function. A considerable fraction of retained introns encode identifiable protein domains or parts thereof 811. In some cases intron retention produces different functional isoforms (EBNA-3 family anigens of the Epstein-Barr virus 12); isoforms with aberrant function (cancerspecific form of cholecystokinin 2 receptor 13); truncated proteins that may be involved in regulation (cold-dependent lipid metabolism in plants 14, nuclear transport of retroviruses 15, autoregulation of splicing 16); non-functional proteins (P-element of Drosophila 17 or rat cytochrome P450 CYP2C11 in stressed liver 18); proteins with unknown function (serine protease kallikrein 1920); or, finally, isoforms with no known functional differences between the variants (hormone urocortin 1 prepropeptide 21, cyclooxygenase 22, D1 dopamine receptor (DR1) interacting protein calcyon 23, mouse homeodomain transcription factor Tgif2 24). At that, intron retention may be conserved in vertebrates, e.g. intron 3 of splicing regulator of the SR family 9G8 16 or species-specific, e.g. intron 2 of Tgif2, present in the mouse gene, but not its human ortholog Tgif2 24.

However, it is likely that many cases of observed intron retention were caused by errors of the splicing machinery. Retained introns are the least conserved type of elementary alternatives 25. Moreover, large scale projects that aim at sequencing of full-length cDNA use normalization procedures to enrich low copy transcripts, and these procedures seem to increase the fraction of underspliced transcripts that retain one or several introns 2627. Traditionally such artifacts in cDNA databases were treated as a nuisance and filtered out in attempts to create "clean" sets of alternative isoforms. We tried to look at introns retained in human cDNA data from another angle, assuming that they capture intermediate states of the splicing process and thus provide a glimpse on the splicing mechanisms.

Another way to look at this mechanism is to analyze consequences of mutations in splice sites. This also has been the subject of several very recent studies. Such mutations have two major possible outcomes: exon skipping and activation of cryptic sites, whereas intron retention is relatively rare 3282930. One of important determinants of the cryptic donor splice site phenotype is the presence of a strong candidate donor splice site in the vicinity of mutated sites 331. Cryptic acceptor splice sites are more frequent in exons than in introns, likely due to depletion of AG dinucleotides upstream of the original acceptor sites 32. There are differences in the distribution of candidate exonic enhancers and silencers between skipped exons and exons with activated cryptic sites 33.

Here we systematically studied aberrant and mutated splicing. Specifically, we compared lengths of affected and adjacent introns and exons, as well strengths of splice sites and distribution of predicted splicing enhancers and silencers in these and adjacent exons and introns. While confirming many earlier predictions, our study also provides a number of new observations that are largely consistent with existing models of the splicing mechanisms.

The overall results of this study seem to agree with the existing biological models. The fact that retained introns are relatively short is consistent with the possibility that such introns are spliced out by the intron definition mechanism, as in this case splicing aberrations should lead to intron retention. When this study was completed, similar observations were made also in 9.

The relative weakness of splicing sites in retained introns and the fact that exons skipped due to mutations of splice sites do not have strong cryptic sites in the immediate vicinity shows that the site scores are a reasonable approximation to site strength and may determine their functionality 33132333940 At that, unlike 3, the relative dearth of cryptic candidate sites in the vicinity of the C-exons was not confined to exclusively to the exons with mutated donor splice sites. On the other hand, we could not confirm the observation that strong acceptor sites are a characteristic of the C-exons with mutated donor sites 31.

In contrast to previous studies that were primarily interested in functional (e.g. conserved) alternative splicing of retained introns 810, we did not enforce possible functionality. One of consequences of that is that the majority of retained introns studied here are unlikely to encode functional proteins, as only 3.3% of them are frame-preserving (this number is close to 4.6% in-frame retained introns observed in Arabidopsis 10). This does not preclude the possible role of such introns in regulation, either on the protein level (e.g. leading to the synthesis of shortened proteins with regulatory function) or on the mRNA level (leading to NMD-inducing isoforms in some specific conditions); some examples of such regulatory mechanisms have been mentioned in the Introduction. However, both the procedure and the obtained results seem to indicate that the majority of retained introns in our study come from underspliced transcripts.

In line with this reasoning, the weakness of sites in retained introns may have two explanations. The retained introns might come from underspliced transcripts (weaker sites imply lower splicing efficiency) or be instances of regulated alternative splicing. Indeed, functional alternative splice sites are weaker than constitutive splice sites 4142. Further, longer introns in general tend to have stronger splice sites; however, the latter trend becomes observable only for bona fide introns longer than 1500 nt 43, and thus should not influence the majority of retained introns studies here.

It has been demonstrated that both human and plant retained introns are more prevalent in the 5'- and especially 3'-untranslated regions, compared to the protein-coding regions of the mRNAs mechanism 810. This has been ascribed to elimination of abnormally spliced mRNAs by the NMD mechanism 44. However, this would not explain the observed prevalence of NMD-inducing retained introns in the 5'-regions. Our results demonstrate monotonic increase in the fraction of mostly retained introns in the 5'-to-3' direction. This is consistent with some degree of co-transciptional splicing (as opposed to simple commitment to splicing with the actual process starting simultaneously for all intron) observed in experiment 45. However, this correlation is not straightforward. Indeed, since we considered only introns bounded on both sides by internal exons, and required that the boundaries of the exon containing the unspliced intron coincided exactly with the boundaries of the corresponding exon-intron-exon chain in the RefSeq mRNA isoform (see Methods), all retained introns considered here are followed by spliced out introns. This means that the observed tendency may not be a simple consequence of completely unspliced 3'-termini.

The observed differences in the density of exonic splicing enhancers in the retained and constitutive introns as well as in the C-exons and S-exons also seem to have a natural biological interpretation. Indeed, a high density of ESE-like sites in an (relatively short) intron may lead to misrecognition of this intron as a part of an exon together with the flanking exons. Similarly, a high density of ESEs in an exon with a mutated site may force the splicing machinery to retain this exon and use a cryptic site, whereas ESSs might provoke skipping the exon. A puzzling observation that candidate enhancers predicted by RESCUE-ESE were more abundant in the constitutively splice introns than in retained ones may be explained by the fact that this method, unlike PESX, is based on the comparison of oligonucleotide frequencies in constitutive and alternative exons and does not control for the distribution of these oligonucleotides in introns 353637. A similar observation was recently made in 33. Another coincidence between our study and 33 is that not all SELEX-based ESEFinder candidate exonic splicing enhancers have different densities in the S-exons and C-exons: in 33, the most pronounced effect was observed for SF2/ASF, whereas in our study a more statistically significant difference was seen for SRp40. In retained introns, the most prevalent candidate splicing enhancers were those for SF2/ASF and SC35, trailed by those for SRp40 and, marginally significant, for SRp55.

Unfortunately, at present it seems impossible to repeat these analyses with intronic splicing enhancers and silencers, since no programs for their recognition are available. A more convoluted, but still plausible explanation may be found for the observed significant difference in the strength of authentic acceptor sites of the C-exons and S-exons: an exon with a weak splice site already contains more splicing enhancers than an exon with strong sites 354647, and thus it is more likely to become a C-exon if the site is disrupted by a mutation.

Thus the analysis of retained introns in human cDNA, exons skipped due to mutations in splice sites and exons with cryptic sites produced results consistent with the intron definition mechanism of splicing of short introns and the model of co-transcriptional splicing. Retained introns tend to be short and contain a higher density of splicing enhancers. Skipped exons contain more candidate splicing enhancers and less silencers, compared to exons with activated cryptic sites. Skipped exons also do not have strong candidate splice sites in the vicinity of mutated ones.

MSG conceived the project. EZK collected and analyzed the data. MSG and EZK wrote the manuscript.