We used an in silico approach based on expressed sequence tags (ESTs) to identify unusual 3' splice sites that are found in pairs of 3' splice variants. ESTs, as first-pass results from high-throughput cDNA sequencing projects, are clearly prone to errors. Therefore, we assumed that single ESTs are insufficient to indicate genuine subtle splice variants since technical artifacts contribute to false positives. We considered a variant as sufficiently evident if it is supported by at least two independent ESTs, and expect the EST variant ratio to serve as an approximation for the natural ratio of splice variants. An additional threshold was applied to the relative abundance of splice variants, since our experimental approach, that is, sequencing of 100 individual RT-PCR clones, had a detection limit. Using a random binomial distribution to model the occurrence of splice variants in the RT-PCR clones, we calculated a diagnostic power of 95% (β error 5%) if a splice variant occurs with at least 3% frequency. It is important to note that we have not inferred anything for the cases that failed the threshold criteria. It is possible that they actually represent natural splice variants; however, the evidence for such cases is weak and the experimental approach did not provide sufficient sensitivity for validation.

We initially aimed to identify unusual 3' splice sites that are found in pairs of 3' splice variants that differ by 3 nucleotides (nt), such as in the GNAS intron 3 splice site tandem 1516. Identification of 3 nt splice variant pairs (Δ3SVPs) was based on 3' splice sites as indicated by spliced alignments of human ESTs 22. After a reduction of false positives performed by a series of filtering steps (Figure 1), we identified 65 'unusual' Δ3SVPs that were supported by high-quality local EST alignments. Of these, 20 meet the requirements that the minor splice variant is supported by at least two ESTs and 3% of the matching ESTs (see considerations below). However, after close inspection and re-sequencing, we identified 6 of the 20 unusual Δ3SVPs as false positives (Additional data file 1), explained by: 3 nt deletion variants due to sequencing errors; mouse ESTs erroneously attributed to human; or alignment artifacts. Another six Δ3SVPs can be explained by SNPs, where the SNP allele corresponding to a NAGNAG splice site motif is not displayed by the human reference genome sequence 21. Strikingly, all the remaining eight Δ3SVPs suggest that TG dinucleotides function as alternative 3' splice sites (Table 1).

Since all the unusual alternative Δ3-nt splice acceptors identified display TG dinucleotides, we investigated their occurrence in a wider scope. Analogously to the screen for Δ3SVPs (Figure 1), we performed a search for alternative TG splice acceptors at larger distances, up to 36 nt from from the canonical splice site. The same filter procedures were applied, and close inspection did not reveal obvious artifacts or explanatory SNP alleles. We identified 26 additional EST-supported splice variant pairs that suggest alternative TG splice acceptors functioning in U2-dependent introns (Table 1).

We sought to screen for unusual splice acceptors using an independent approach in order to cross-validate our initial findings and to make a link to previous studies that were primarily based on curated transcript data 59. An analysis of RefSeq-to-genome alignments identified 122 putative introns with terminal TG dinucleotides (120 unique genomic sites) out of 228,925 total introns (171,605 unique genomic sites). Of these, 39 introns have a canonical GT donor dinucleotide (Additional data file 1). A previous study, performing a similar screening approach for unusual splice sites using curated transcript data, showed high enrichment of annotation artifacts 9. Therefore, we checked the identified TG acceptor cases thoroughly. In fact, cases failed this quality check for several reasons: known SNPs masking existing canonical AG splice sites; RefSeqs lacking transcript evidence; and misleading RefSeq-to-genome alignments. Since the overall false positive rate seemed very high, we additionally required independent transcript entries (mRNA or EST) to support the unusual splice site. In summary, 9 of the 39 RefSeqs showed robust support for unusual TG acceptor sites (Table 1, entries in bold), and 6 of these overlap with cases obtained from the EST screening approach while the others are exclusively identified by the RefSeq-based approach (SH3D19, BAT3, CACNA1A). Intriguingly, these three EST-independent RefSeq-supported cases all comprise 'alternative' splice sites, although this was not a screening criterion. Taking into account that about 1% of all introns have alternative 3' termini 23, this strongly indicates that TG splice acceptors are functionally linked to nearby AG splice sites and cannot function in a constitutive manner (P = 0.000001, binomial test). Altogether, the two screens identified 37 introns with 39 alternative TG splice acceptors (Table 1).

Since six putatively unusual 3' splice sites can be explained by SNP-affected NAGNAG acceptors (which were filtered; Figure 1), we asked whether undiscovered SNPs, or even inaccuracies in the available human genome sequence, may explain some of the remaining candidates. The genomic sequence of the splice site regions of GNAS and CACNA1A had been experimentally verified by others 1624. For 10 other genes (listed in Table 2), we analyzed PCR products obtained from genomic DNA, pooled from 100 individuals, for sequence variations. The re-sequenced genomic regions were in perfect agreement with the available genome sequence, negating the possibility that unusual splice sites are trivial sequencing errors (data not shown). Moreover, we identified no SNP alleles that confer explanatory AG splice sites on any of the observed unusual splice variants, demonstrating that the TG splice sites are real and genetically invariant.

To verify the existence of TG-derived splice variants, we performed RT-PCR experiments designed to yield amplicons that cover the exon-exon junctions under consideration. Cloning of the PCR products and sequencing allowed us to detect splice variants. Subclassification and counting of clones gave measures of splice variant ratios. This way, the alternative splicing pattern was reproduced and quantified for 11 out of 12 analyzed cases (Table 2). Generally, the splice ratios obtained from clone counting agree well with the EST data. The observed deviations can be explained by significant fluctuations depending on the analyzed tissue (C21orf63, BRUNOL4, and CNBP in Table 2). The splice variant validation failed for FBXO17, a gene for which 4 out of 29 ESTs had suggested a TG-derived splice variant. All supporting ESTs originated from the same EST library, NIH_MGC_100, derived from a hepatocellular carcinoma. A peculiarity of the source material, either the NIH_MGC_100 cell line or the single-individual liver sample used for our RT-PCR experiments, may be the reason for the inconsistent results concerning this putative splice variant. This example illustrates that at least two ESTs from independent sources are required to indicate a natural splice variant with high reliability. Overall, the success rate of the validation experiments was high (92%), and extrapolated to the 25 non-tested cases, about 2 false positives are expected.

According to the results of the PCR cloning approach, BRUNOL4 displayed remarkable tissue-specific splice ratios. The TG-derived splice variant was not detected in lung cDNA whereas the same variant constituted 20% of brain BRUNOL4 transcripts (Table 2, Figure 2a). So we asked if splice ratios of TG-derived and AG-derived variants generally show tissue-specific differences. We analyzed the splice variant ratios more extensively in other genes using pyrosequencing, a method that allows accurate and cost-effective quantification of mixtures of polymorphic DNA populations 2526. ARS2 and CNBP, both having a ubiquitous expression profile, show tissue-dependent fluctuations in splice ratios (55-65% and 20-40%, respectively; Figure 2). While these differences are numerically significant in each of these genes (α = 0.01, ANOVA), their biological relevance is debatable. We conclude that splice ratios of TG-AG tandems are tissue-specific for particular introns but are rather stable for others.

Since splicing at TG sites occurs in a very small number of introns, one might argue that these represent 'accidental' events attributable to spliceosome dysfunction. To address this question, we first analyzed the conservation of splice sites in homologous introns as an indication of alternative splice variants being under purifying selection. Out of 36 introns with 3' TG splice sites (37 minus the false-positive FBXO17 intron), 26 (72%) are conserved between the human and mouse genomes. In 14 of these cases (39% of the total), mouse ESTs indicate homologous TG-derived splice variants. For comparison, this rate is three- to four-fold higher than that of alternative exons found in both human and mouse 272829. In some cases, EST evidence for orthologous TG-derived splice variants even exists for distantly related species, such as chicken (CNBP, BRUNOL4, RYK), and frog or fish (BRUNOL4, RYK, FBXL10). An outstanding example of conserved intron sequence and homologous splice variants is intron7 of the RYK gene (Figure 3). The ratio of 3' splice variants is remarkably similar between human and chicken, as can be inferred from the available EST data (EST ratios of 24:6 and 5:1 for human and chicken, respectively). In general, it should be noted that homologous splice variants may remain undetected due to the limited depth of EST coverage 2327.

Independently, we analyzed intron sequence conservation as an indication of the functional relevance of alternative splicing 272829. A data set of human-mouse orthologous intron-exon boundaries was used to determine the degree of conservation within a 50 nt intron sequence upstream of the splice acceptor, or acceptor tandem. Intronic flanks of TG splice sites show an average sequence similarity of 74%, whereas flanks of AG splice sites within canonical (AG-only) introns are 65% similar on average (P << 0.00001, permutation test). A plot of flanking sequence conservation against the abundance of the TG-derived splice variant (Figure 4) shows that these two measures are positively correlated. Introns that give rise to less than 10% TG-derived splice variants have an average human-mouse intron sequence identity of 64%, indistinguishable from canonical introns. In contrast, introns with a TG-derived splice variant making up more than 10% of the transcripts show an average sequence identity as high as 80%. This parallels a previous finding that the abundance of splice variants correlates with sequence conservation of alternative exons 27. Consistently, high intron sequence conservation is strongly correlated with conservation of the TG splice site (13 of the 14 cases with gene labels in Figure 4). Our results indicate that splicing at TG acceptors may arise from neutral evolution, presumably showing low splicing efficiency. However, efficiently spliced TG 3' splice sites seem to evolve and to be maintained by evolutionary selection.

We analyzed the context properties of TG 3' splice sites in order to find an explanation for these rare exceptions to the GT-AG rule. With regard to the gene structure context, TG-splicing introns are indistinguishable from canonical introns in several respects: length of the affected as well as the downstream introns, and length of the upstream and downstream exons (data not shown). TG 3' splice sites are significantly often found in the first intron of the gene; 8 of 36 (22%) TG-AG introns compared to 11% of other introns (P = 0.02, Fisher's Exact). This bias is also found for AG-AG splice site tandems and is certainly due to neutral evolution of introns located in the 5' untranslated transcript region.

TG 3' splice sites were exclusively found within a context of alternative splice site choice (Table 1). This is clearly significant for results from the RefSeq-based screening procedure, which are unbiased with respect to constitutive or alternative transcripts. Taking into account that about 1% of all introns have alternative 3' termini 23, it strongly indicates that TG 3' splice sites are functionally linked to nearby AG splice sites and cannot function in a constitutive manner. This conclusion is further supported by studies of human AG>TG 3' splice site mutants 303132, which always resulted in activation of neighboring AG splice sites, but not splicing at TG. Furthermore, we observed that TG-AG tandems display a splice site distance restriction with a limit of 28 nt, which is not seen for AG-AG tandems (Table 1, Additional data file 1). Thus, splicing of a 3' TG does not depend only on an additional AG splice site, this dependency also seems to pose a constraint on TG-AG splice site distance. The observed distance limit corresponds well with the distance between the branch site and the 3' splice site, which is typically 20-40 nt 33.

Splice site strength was scored using the maximum entropy method 'maxent' of Yeo and Burge 34. The 5' splice site scores are indistinguishable between TG-AG introns and canonical introns. On the other hand, the AG 3' splice sites in TG-AG tandems score significantly lower than canonical introns (6.1 ± 3.5 versus 8.5 ± 2.8, respectively; P = 0.00001, Student's t-test). The 3' splice site score remains significantly small if low-scoring outliers are excluded from the analysis (6.7). Interestingly, the sequence context of TG 3' splice sites is very similar to canonical AG splice sites in that it shows a preference for pyrimidines at position -3 and preference for purines at position +1 (Additional data file 1). However, TG 3' splice sites changed to AG yield an average score of 6.6, again significantly lower than that of canonical AG splice sites (P < 0.00001, Student's t-test). This disfavors the simple explanation that TG and AG splice sites compete for the same recognizing factors and the neighboring nucleotide composition (that is, the feature scored by maxent) alone acts to direct splice site choice towards TGs. Finally, it remains questionable if maxent, trained on canonical AG 3' splice sites, has any predictive power for the functionality of TG splice sites.

The fraction of TGs functioning as splice acceptors is extremely small, about 0.01% of candidate motifs. Thus, cis-regulatory elements must play a crucial role in the definition of TG splice acceptors. From the ratio of functional/non-functional TG-AG tandems, in comparison with AG-AG tandems, we estimate that at least 6 nt of cis-regulatory sequence information is required to promote splice site usage of 3' TGs (Additional data file 1). This is in agreement with 5-20 unchanged nucleotides in excess over the average intron mutation rate, found in about half of the TG-splicing introns, that is, those considered to be subject to purifying selection for splicing of 3' TGs. However, we failed to identify specific regulatory motifs (data not shown). This may be due to the dispersed arrangement of cis-regulatory elements, or the contextual cooperation of diverse elements. Due to the relatively small sample size for TG 3' splice sites, available methods for motif discovery have limited detection power.

From the UCSC Genome Browser site 22 we obtained spliced alignments of human ESTs (file all_est.txt, released 2005-07-14) and of human RefSeq transcripts (refGene.txt, 2005-07-23) 49, as well as a compilation of all human EST sequences (est.fa, 2005-11-26). First, we sampled EST-supported 3' splice sites to identify 3-nt splice variant pairs (Δ3SVPs). In parallel, we identified ESTs that were mapped to multiple genome locations, indicative of paralogous gene loci including pseudogenes. We discarded those Δ3SVPs whose EST support for the minor splice variant did not exceed the number of these ambiguously mapped ESTs. Furthermore, we retained only those Δ3SVPs that have at least one splice site corresponding to a RefSeq transcript, according to the RefSeq-to-genome alignment. Then, we separated cases that involve the dinucleotide AG at both 3' splice sites, that is, NAGNAG tandem splice acceptors, as well as U12-dependent introns, identified by their characteristic donor site and branch-point motifs 6. The remaining Δ3SVPs were considered 'unusual' since they comprise at least one non-AG splice acceptor in a U2-spliceosomal intron. The splice variants of these Δ3SVPs were validated and quantified by a WU-BLASTN search of 60-nt sequence windows around the resulting exon-exon junctions against all human ESTs, using parameters W = 13, N = -8, nogap S = 180, hspmax = 1. BLAST matches were considered valid if perfect sequence identity was found in a 12-nt window around the exon-exon junctions 37. Finally, Δ3SVPs were considered highly reliable if the minor 3' splice site was found in at least two ESTs and was used in at least 3% of the covering ESTs. A screen for splice variant pairs for distances of 4-36 nt was performed analogously, restricting the search to tandems of TG-AG splice sites.

For validation of splice variants, nested PCR was performed using 1 ng cDNA templates from the Human Multiple Tissue cDNA PanelsI and II (Clontech, Mountain View, CA, USA). For a given gene, suitable tissues were determined from expression data obtained from the Stanford SOURCE database 50. However, pooled leukocyte cDNA from 200 individuals was preferably chosen in order to obtain comparable results. Verification of the genomic sequence and an analysis of potential polymorphisms were done by nested PCR using 200 ng of pooled genomic DNA from 100 Caucasian individuals (Roche, Mannheim, Germany) as template. Primers were obtained from Metabion (Martinsried, Germany) (Additional data file 1). Reactions were set up with PuReTaq Ready-To-Go PCR beads (GE Healthcare, Munich, Germany) and 10 pmol primer in 25 μl total volume, according to the manufacturer's instructions. A typical thermocycle protocol was 3 minutes initial denaturation at 94°C, followed by 25 cycles of 1 minute denaturation at 94°C, 1 minute annealing at 53-55°C, 1 minute extension at 72°C, and a final 10 minute extension step at 72°C. In the second round of nested PCR, 1 μl of the first-round product was amplified for 30 cycles. For cloning, PCR products were separated on agarose, DNA was extracted applying the Millipore (Billerica, MA, USA) Montage Gel Extraction kit, followed by ethanol precipitation. Isolated fragments were cloned in pCR2.1-TOPO (Invitrogen, Karlsruhe, Germany), and cloned DNA was Sanger sequenced using M13 standard reverse primer (17-mer).

Templates for pyrosequencing were generated using universal biotinylated primers 51. RT-PCR amplicons of the exon-exon junctions were ligated into pCR2.1-TOPO (Invitrogen) according to the supplier's recommendations and subsequently re-amplified with all four possible combinations of 5'-biotinylated M13 standard primers (17-mers) and unlabeled insert-specific primers (Additional data file 1). The latter also served to prime the pyrosequencing reaction. Biotin-labeling of DNA, single strand preparation and sequencing were performed as described 51.

A data set of orthologous intron-exon boundaries was constructed automatically to obtain sufficient data (especially reference data) to test for evolutionary constraints on intron flanking sequence. Sets of human (data as described for the splice site screen) and mouse transcript annotations (UCSC genome assembly mm7, RefSeq-to-genome alignment 2006-05-21) were processed as described earlier (supplementary methods in 20). For 97,107 unambiguous orthologous pairs (57% of unique human intron-exon boundaries, including 23 of 36 TG-AG splice site tandems), 100 nt flanking intron sequences were aligned using CLUSTALW 52, using the optimized parameter -gapopen = 2.2. The degree of conservation was determined for 50 nt of the human intron sequence upstream of the splice site (tandem), giving a score of 1 for an identical aligned nucleotide in mouse, a score of 0 for a mismatch, and a penalty of -1 for inserted mouse sequence. Since a histogram of sequence conservation in canonical introns showed a non-normal distribution, statistical testing was performed using a permutation test. Intron samples of given size were simulated by random drawings from the intron data set, and the average sequence identity was calculated, repeating the sampling procedure 10,000 times.

Where automated processing failed (13 of 36 TG-AG splice site tandems), orthologous intron-exon boundaries were retrieved using the UCSC genome browser. These cases were not used for the statistical analysis since these represent a likely biased subset with regard to sequence conservation, and an appropriate large data set for comparison is not available.