Two computer programs, GeneSeqer 34 and GMAP 35 , produced largely similar results for the alignment of EST sequences to their native genomes for the Mt, Lj, At, and Os data sets. To reduce the likelihood of alignment artifacts as a result of ambiguities, only the commonly predicted alignments from the two programs were used in further analyses. Moreover, highly stringent criteria (>95% sequence identity, >80% transcript coverage) were used to limit the possibility of transcript mapping to non-cognate, diverged locations in the incompletely sequenced genomes. Approximately one half and one third of the species-specific EST sets could be aligned to the current Mt and Lj genome sequences, respectively, roughly reflecting the coverage of the whole genomes by their current sequence assemblies. For Lj, ~15% of the transcript sequences were mapped to finished (phase 3) BAC/TACs. Unless stated otherwise, our analyses for Lj were based solely on this subset. As shown in Table 1, a total of 11,516 and 3,298 genes/transcription units (TU, as defined in METHODS) were identified in Mt and Lj, respectively, with 74% and 57% of them having multiple EST support. The average number of ESTs per gene/TU was 10 and 7 in Mt and Lj, respectively, compared with 26 and 30 in At and Os.

We compared intron/exon features revealed by EST alignments in the four species. The intron size distribution was quite similar in Mt and Lj, with a mean intron size around 460–470 nt and median approximately 220 nt in both species. Legume introns are therefore significantly longer than in At (mean 171 nt, median 101 nt) and slightly longer than Os introns (mean 438 nt, median 164 nt). As shown in Figure 1A, the intron size distributions have a peak near 90 nt in all four species. Mt and Lj have fewer introns shorter than 150 nt but more introns longer than 200 nt compared with At and Os. At introns are clearly the shortest of the four plants. Fewer than 1% of introns are longer than 1,000 nt in At, while this number is over 10% in the other plant species. Exon size tends to be similar among the four plant species, with legume exons slightly shorter than At and Os exons. In Mt and Lj, the mean internal exon sizes are 140 and 127 nt, respectively, with the median sizes about 108 nt and 100 nt. At and Os have internal exons with a mean of 164 nt and 175 nt and a median of 113 nt and 114 nt. Figure 1B shows that the size distributions of exons in Mt, At and Os all display a peak at around 80 nt. Lj data is less consistent due to its small sample size. In contrast to introns, the frequency of exons smaller than 150 nt is higher in Mt and Lj than in At and Os, while the frequency of exons longer than 200 nt is lower in legumes. Overall, legumes have longer introns but slightly shorter exons than At and Os. Generally speaking, plant introns are longer than exons. More than 40% of introns in Mt, Lj and Os are longer than 300 nt, while less than 10% exons are so large.

As noted previously 1 36 , the GC-content of introns and exons is ~5% lower in At than in Os. The GC-content of legume introns and exons is very similar to that of At, although Mt has slightly lower GC-content than either At or Lj in both intronic and exonic regions (see Additional file 1, Supplementary Table S1 and Supplementary Figure S2). G-content and A-content are similar in all species including Os, although Os introns are relatively more C-rich and less U-rich. There is more variation in the distribution of U-(T-) and A- content than in G- or C-content in all species (see Additional file 1, Supplementary Figure S3). The difference in GC-content between introns and exons is about 10% in all four species, with Mt showing the largest difference of 11.7% and Os showing the smallest, 9.6% (see Additional file 1, Supplementary Table S1).

Previous studies revealed that approximately 20% of expressed genes are alternatively spliced in At and Os, with half of the AS events being intron retention (IntronR) 1 . When we re-examined AS frequency in At and Os for this study, we also found a frequency of around 20%. However the total number of transcript sequences increased 80%-200% due to the increased sizes of the EST data sets in these species. In the case of Mt and Lj, the number of ESTs available for analysis were much lower. Consistently, the fraction alternatively spliced genes observed was much lower, just 9.6% in Mt and 2.8% in Lj (Table 2). Examples of alternatively spliced genes in Mt are shown in Additional file 1, Supplementary Figure S1. All the AS data are deposited and viewable at the ASIP site 37 .

To compare the frequency of alternative splicing between different species, earlier studies relied on 10 randomly selected ESTs per gene as a basis for estimating AS frequency 4 . Here, only a small fraction (10–20%) of legume genes were covered by 10 or more ESTs, so this approach was not practical. Instead, we plotted the AS frequency for all groups of genes with similar EST coverage in different species, as shown in Figure 2. Mt categories with fewer than 80 genes total were removed to reduce noise due to small sample size, and Lj data are not included at all, as sample size was uniformly too small. When analyzed in this way, the fractions of alternatively spliced genes are similar regardless of species for nearly all size classes. For genes with four ESTs (the median EST number per gene in Mt), the observed AS frequency is 6–12% in Mt, At, and Os alike. For genes with nine to 11 ESTs (the median EST number per gene in Os and At), 15–23% are alternatively spliced. In general, the fraction of alternatively spliced genes keeps increasing with increasing transcript coverage, eventually reaching 66% in Os and 46% in At for genes with hundreds of ESTs, a levels similar to those observed in mammals 38 39 . Interestingly, the AS level in Os is consistently over 10% higher than in At in genes with more than 40 supporting ESTs.

As shown in Table 2, the proportions of different AS types are similar in Mt, At and Os. (Lj data are also listed but are not included in the analysis as only ~100 AS events were identified). More than half of AS events in plants are IntronR, 6–11% are ExonS, and the remaining 30–40% involve different splice sites (AltD/A/P). These numbers are quite similar to those observed previously 1 . Mt has a slightly lower ratio of IntronR (51%) and a higher ratio of AltD (13%) compared with At and Os. Different levels of EST coverage have little effect on the composition of AS events. As shown in Additional file 1 (Supplementary Figure S4), the ratios of different AS types remain largely constant across all EST levels, particularly in At and Os. IntronR is the most abundant at all levels, with a relatively lower ratio in Mt. The ExonS ratio is consistently lower in At than in Os (and Mt), while the AltA ratio is higher.

To minimizes false AS events caused by sequencing errors or contaminations in the EST collection, we repeated the above analysis for the subset of AS events that are supported by at least two transcript sequences 40 . As shown in Figure 3, the ratio of IntronR decreased ~5% in all plants in this subset. Mt has the lowest ratio of IntronR (45%), 6–7% lower than in At and Os. The ratio of ExonS remains unchanged compared with the full data set. In Mt and Os, 10–11% AS events are ExonS compared to 7% in At. The AltD ratio in Mt increased significantly to 21% in the subset, nearly double the ratio in At and Os. In At, the AltA ratio is ~30% compared to 23% in Mt and Os. Similar tendencies were observed for subset data with even more transcripts supporting each isoform. Both the full and subset data indicate that Mt has a lower ratio of IntronR and a higher ratio of AltD, and that At has a lower ratio of ExonS but a higher ratio of AltA.

Even "reliable" AS events (as defined above) may not necessarily be functional. Because conservation is usually a good indicator of function, we deployed a cross-species approach similar to large-scale methods used previously in mammals to identify functional AS events 7 9 . All available EST sequences from Lj, Gm, and other legume species were aligned against Mt BACs. One concern with the cross-species approaches has been a potentially high error rate 7 . Here, even using an identity cutoff as high as 80%, hundreds of AS events were identified from either GeneSeqer or GMAP alignments alone, with approximately 40% of events consistent between the programs. Our analysis used only common events identified by the two programs to reduce false positive events from alignment errors. As shown in Table 3, 10–20% of the non-Mt legume transcript sequences could be mapped to Mt BACs and clustered to a total of 7,896 non-redundant genes, 81% of which have also Mt EST support. Approximately 70% of the introns identified from cross-species EST alignments were consistent with Mt EST supported introns. The gene structures derived from cross-species ESTs and Mt ESTs alignments were mostly consistent, demonstrating the value of cross-species ESTs in genome annotation 10 . In this analysis, a total of 307 Mt genes (3.9%) were found to be alternatively spliced, with 248 genes having no evidence of AS from Mt ESTs alone. If these novel AS events are included, the estimated frequency of Mt alternatively spliced gene increases from 9.6% to 10.4%. Interestingly, many more AS events were identified from soybean ESTs than from Lj ESTs, despite the similar evolutionary distance between Mt-Gm versus Mt-Lj. At and Os EST sequences were also applied in a comparable cross-species analysis, but only 1% of them could be mapped using the same criteria. No reliable AS events were deduced from At and Os transcript sequences.

Altogether, 367 cross-species AS events were identified from legume cross-species EST alignment, including 35.7% IntronR, 16.9% ExonS, 16.1% AltD, 29.1% AltA, and 2.2% AltP (Table 4). Compared with AS events identified using Mt ESTs alone, the cross-species AS events display a relatively lower ratio of IntronR and higher ratios of ExonS, AltD, and AltA. As most of the cross-species AS events are likely conserved between Mt and the native species of the EST, the ratio of each AS type in cross-species AS events could be interpreted to represent the ratio of functional AS events. However, the ratio of IntronR could have been underestimated by cross-species EST alignments because intron sequences are not as well-conserved as exons, even in closely related species. Thus, some cross-species ESTs retaining introns from their native species might have been filtered by the 80% identity cutoff. The location and outcome of cross-species AS events and same-species AS events are compared in Additional file 1 (Supplementary Table S2).

Approximately 90% of cross-species AS events are located in open reading frames (ORFs), much higher than the fraction (70–75%) in same-species AS events. There seem to be more cross-species and same-species AS events in the 5'-UTR than in the 3'-UTR (data not shown and 1 ). For AS events in ORFs, the fractions of translation-readthrough events, where some amino acids are added to or removed from the protein without changing the reading frame, are similar (20–24%) in cross-species and same-species events. AltA has the highest translation-readthrough ratio (35–40%), and IntronR has the lowest (2–10%). Intriguingly, the ratio of AS events producing substrates for nonsense-mediated decay (NMD) 41 is higher in cross-species AS events than in same-species AS events. Nearly half of the cross-species AS events produce NMD substrates, compared with 30–40% in same-species AS events.

To identify AS events with direct evidence of conservation in multiple species, two approaches were employed: (1) Align all legume ESTs to Lj TACs to identify conserved AS events predicted by the same ESTs between Mt and Lj; (2) Identify conserved AS events in Mt with EST evidence from multiple legume species, all showing the same AS pattern. A total of 242 AS events conserved between Mt and Lj were identified through method (1), including 92 (38.0%) IntronR, 26 (10.7%) ExonS, 78 (32.2%) AltA, 41 (17.0%) AltD, and 5 (2.1%) AltP events. These AS events are viewable at the ASIP website. Method (2) identified 22 completely conserved AS events in Mt (see Additional file 1, Supplementary Table S3). Nine of the 22 genes also have At and/or Os close homologs sharing the same AS pattern. For instance, Mt hypothetical protein AC156627_1 has both soybean and Mt ESTs support for an AltA event in the first ORF intron, whereby an isoform utilizes an alternative acceptor site 5-nt upstream (AACAG) of the constitutive acceptor site (AGCAG), producing a substrate possibly subject to NMD. At homologs (At5g25360.1 and At1g15350.1) and Os homolog (LOC_Os02g10720) both have exactly the same AS pattern, including the alternative acceptor sites. This gene seems to be plant-specific, as non-plant homologs can not be identified. Another example of completely conserved AS events is the Mt AP2 domain containing protein AC151460_3, where the 3'-UTR intron can be retained. One At homolog and three Os homologs also have the same intron retained. There are also some AS events conserved in legumes but not observed in At and Os. One example is AC124951_11, a highly expressed carbonic anhydrase gene with the 3'-UTR intron alternatively spliced (AltD) in legumes species. The AltD event is conserved in all legume species (Mt, Lj, Gm, and others), but not in At and Os even though hundreds of ESTs exist, indicating that this AS event is probably legume-specific.

One example of a completely conserved ExonS event occurs in an enoyl-CoA hydratase/isomerase gene (Mt: AC145449_47). As shown in Figure 4A, the IMGAG-annotated gene structure for AC145449_47 contains 11 exons, each with strong EST support. Exon3 (65 nt) and Exon4 (53 nt) are mutually exclusive. In one isoform, Exon3 is retained and Exon4 is skipped (Mt: 7206545, 90656179; Lj: 45578881; Lupine: 27458685). In another isoform, Exon4 is retained with Exon3 skipped (Mt: 7567285, 11904359, 13596489, 33106093; Lj: 7719575). The two mRNA isoforms therefore encode two proteins (418 aa and 414 aa) differing slightly in their predicted Enoyl-CoA hydratase domain (ECH, pfam00378). No isoform contains both exons, while it is possible to skip both (Mt: 83667352). Two genes in At (At4g13360 and At3g24360), one gene in Os (LOC_Os06g39344) and one in Lj (LjTC_2465, AP006370.1: 88858–94512) are the closest homologs to AC145449_47. Exactly the same AS pattern was observed in all the homologous genes except for At4g13360, where the 65-nt exon (Exon3) was retained constitutively and no trace of the 53-nt exon can be found in the corresponding region (Figure 4C–E). Sequence comparison revealed several nucleotide bases in degenerate codons conserved in all four species (Figure 4B). These bases may contribute to the recognition of (or skipping) the exon.

In this study, alignment of current EST and genomic sequences revealed that ~10% of expressed genes are alternatively spliced in Mt compared with 20% in At and Os. This difference is mainly due to the lower EST coverage found in Mt. We demonstrated that the AS frequencies in the three plants are essentially similar when adjusted for genes having comparable EST numbers. This conclusion is different from the conclusion drawn in a recent study based on EST pairs gapped alignments, in which a greater degree of variation was observed for different plant species 3 . Interpretation of EST-only data can be confounded by extensive gene duplication events. With more plant genome sequences becoming available, it should soon be possible to more precisely address the intriguing questions concerning the extent and evolution of AS in plants.

Alternatively spliced isoforms are usually in low abundance, the chance of capturing them in a small EST collection is low, making it difficult to estimate AS frequencies accurately. Supposing a functional event has certain percentage p of transcripts alternatively spliced, the probability of observing an AS event with n ESTs covering the alternative splice site is 1 - (1 - p) ⁿ . For example, if an alternatively spliced isoform were generated p = 10% of the time, n = 10 transcript sequences would give a 65% probability of observing this event, and 22 transcript sequences would be required to have >90% probability of observing the event. Our results show that the AS frequency for genes with small numbers of ESTs are similar in Mt, At, and Os, suggesting that they all have similar levels of functional AS events.

In cases where AS isoforms are even lower in abundance, greater numbers of transcripts would be clearly necessary to detect the event. Nevertheless, Os seems to have a higher frequency of AS in genes with >30 ESTs than either Mt or At. Focusing on genes with >40 ESTs only, the AS frequency in Os is consistently (>10%) higher than in At. For this analysis, we did not include transcripts from Os subspecies indica in order to eliminate the possibility that the higher AS frequency is falsely caused by cross-subspecies ESTs. In any case, the error rates from EST sequencing or genome contamination are probably similar in all three plants. Consequently, Os does seem to have higher levels of low-abundance AS events than At (or Mt). Some of the low-abundance events may be splicing errors captured in EST libraries constructed from plant tissues under various growth conditions, so the higher level of low-abundance AS events in Os could indicate higher error rates for the Os spliceosome.

Not surprisingly, observed AS frequency is highly correlated with EST numbers in all three plants. Highly expressed genes (genes with large numbers of ESTs) are more likely to be detected as alternatively spliced. Over 60% and 40% genes with more than 500 ESTs are alternatively spliced in Os and At, respectively. This is comparable to the level in human 42 . Half of human genes are alternatively spliced by the criterion that AS isoforms occurs in at least 1% of the observed transcripts, but only 20% of human genes are alternatively spliced if the required abundance level is increased to >10% 42 . This frequency is notably similar to the frequency in plants under the same abundance level, suggesting that the frequency of regulated AS events in plants may not be significantly lower than in mammals.

A clear difference between AS in plants and mammals is the predominance of IntronR in plants and ExonS in mammals. Both model legumes, Mt and Lj, have 40–50% of AS events as IntronR, a level noticeably lower than in At and Os, but still much higher than in mammals. Similar to the situation in At and Os 1 , introns shorter than 70 nt are more likely to be retained in legumes (data not shown). The spliceosome is a large dynamic RNA-protein complex involving hundreds of proteins. If an intron is too small, the assembly and structure transformation of spliceosome will be constrained and may lead to inefficient splicing and IntronR 1 . As the size of introns is considerably larger in Mt and Lj, fewer introns will be retained due to steric hindrance, possibly leading to a lower frequency of IntronR in legumes. These data also suggest that some AS events may be splicing errors. As we proposed in 1 , the most common splicing error in plants is probably a failure to recognize and splice out introns, so IntronR should be the most common AS type. In mammals, where introns are defined through an exon recognition mechanism, a failure to recognize some exons, and therefore skip them, is likely the most common error. Consequently, ExonS is the most common AS type in human.

Observed AS events are a mixture of functional AS events and splicing errors. Other types of error, such as sequencing errors, genome contamination, and alignment errors, will also contribute to the predicted level of AS events. Two alignment programs (GeneSeqer and GMAP) were applied and only common AS events were used in this study to minimize alignment errors. Genome contamination could be minimized by elimination of ESTs retaining all predicted introns. Distinguishing functional AS events from splicing errors, however, is not an easy task. We attempted to achieve this goal by two methods. First, we selected AS events with each isoform supported by multiple transcripts. As splicing errors are expected to occur at low frequency, the chances they will be captured in two distinct transcripts are low. In this data set, the frequency of IntronR is slightly lower, but still the highest among the five AS types, indicating that IntronR is indeed the most abundant regulated AS result. The second method is to look for conserved AS events through cross-species EST comparison and orthologous gene comparison. A few AS events were completely conserved in Mt, Lj, At and Os.

Functional AS events, however, may not always be conserved. As a dynamic process, splicing requires hundreds of proteins as well as some snRNAs to function accurately 14 . Mutations in both trans- and cis-elements on target genes will impact splicing patterns. Depending on when the mutation and fixation event occurs, functional AS events can be shared among closely related species or be lineage-specific. The AltD event in 3'-UTR of the highly expressed carbonic anhydrase gene (AC124951_11) may be a good example shared by legume species. Lineage-specific functional AS events are difficult to define from EST data alone.

As more plant genomes and ESTs are being sequenced, more AS events will be identified in the future. It is important to have a centralized place to store and compare all AS data. In animal systems, a comprehensive database, ASAP 43 includes AS data from 16 sequenced animals, which makes a comparison across different animal species straightforward. Such a database is also needed in plants, as the study of splicing signals and alternative splicing are just starting. The AS data identified in this study have been deposited in the ASIP database at PlantGDB 37 , where previous AS data are stored and can be easily compared 1 . Moreover, a database collecting genes related to splicing in At, animals and yeast is available through the SRGD database at PlantGDB 14 44 . In the future, the database will be expanded to Os and other sequenced plant genomes including Mt, Lj and poplar. The analysis programs and plant genome browsers available at PlantGDB should facilitate the deep mining of AS data in plants. A core data set in which the AS events are conserved in all sequenced plants will be extremely useful for understanding the function of AS events, as well as the signals and regulation of this important and intriguing phenomenon.

The Medicago Genome Sequence Consortium (MGSC) release 1.0, consisting of the 1,826 BACs analyzed in this study, were downloaded from Medicago genome sequencing project website 45 . The assembly comprises a total of 186.2 Mb of non-redundant genome sequence, an estimated 38–47% of the entire genome and 55–58% of total gene space 46 . All other sequence data sets used in this study were current as of July 17, 2006, the cutoff date for BACs incorporated into the Mt1.0 genome assembly. For Lotus japonicus, 1,394 BAC/TACs were downloaded from the NCBI 47 nucleotide database using the query "txid34305 [ORGN:noexp] AND HTG [KYWD]". Arabidopsis genome sequences and gene annotation (TAIR release 6.0) were downloaded from the GenBank FTP site 48 , and rice genome sequences and gene annotation (TIGR release 4.0) were downloaded from the TIGR FTP site 49 .

All EST sequences (including full-length cDNAs) were retrieved from GenBank nucleotide database. Sets of 225,920 Mt and 150,855 Lj transcript sequences were collected using the queries (txid3880 [ORGN] AND "biomol mrna" [PROP]) and (txid34305 [ORGN] AND "biomol mrna" [PROP]), respectively. Soybean transcript sequences (359,834) were retrieved using the query (txid3847 [ORGN] AND "biomol mrna" [PROP]), and 127,684 transcript sequences from all other legumes were retrieved by using the query (txid3803 [ORGN:exp] NOT txid3880 [ORGN] NOT txid34305 [ORGN] NOT txid3847 [ORGN] AND "biomol mrna" [PROP]). For At, 691,516 transcript sequences were retrieved using the query (txid3702 [ORGN] AND "biomol mrna" [PROP] AND srcdb_ddbj/embl/genbank [PROP]). For Os, 1,009,574 ESTs from the japonica cultivar-group were retrieved using query (txid39947 [ORGN] AND "biomol mrna" [PROP] AND srcdb_ddbj/embl/genbank [PROP]). We intentionally excluded transcript sequences from the indica cultivar-group to reduce possible false positive alignments caused by differences between the two Os cultivar-groups.

The legume transcript sequences were mapped to the Mt and Lj BAC sets using the two computer programs GeneSeqer 34 and GMAP 35 . The splice site models for GeneSeqer were set to Medicago-specific parameters using the program option "-s Medicago". Default parameters were used for all other options. Default alignment parameters were used for GMAP. For At and Os, only GMAP alignments were performed locally, and GeneSeqer alignments derived from a larger data set were downloaded from PlantGDB 50 .

GMAP and GeneSeqer output alignment files were processed by a pipeline (ASpipe1.0, available through SourceForge 51 ) developed from Perl and shell scripts used in a previous study 1 . ASpipe extracts coordinates and scores for high-quality intron/exon/alignments from the original program outputs and stores them in MySQL5.0 databases. For same-species EST alignments, the criteria for high-quality alignments were >95% sequence identity and >80% coverage (defined as the portion the transcript sequence aligned to the genomic sequence). The high identity (95%) cutoff minimizes false mapping of transcript sequences to incomplete genomes. For cross-species transcript alignments, the identity cutoff was decreased to 80%, which selects reliable alignments from divergent transcript sequences. Redundant EST alignments in Mt were removed by comparison with the non-redundant gene list provided for Mt1.0 33 . Exons mapped with >95% and >80% sequence identity were considered as reliably identified exons for same-species and cross-species mappings, respectively. Introns with reliable neighboring exons on both ends were considered as reliably identified introns. A transcription unit (TU) was defined as a consecutive genomic region where transcript sequences were mapped and clustered. Annotated gene models may contain multiple TUs. For Mt, At and Os, annotated genes were used as the base for analysis. For Lj, where no gene annotation is available, TUs were the base for analysis.

The coordinates of reliable introns and exons were compared in a pairwise fashion in order to identify candidates for AS events. For intron/intron comparison, if two introns had the same 3'-end but a different 5'-end, this event was classified as AltD. If two introns differed only in the 3'-ends, this event was classified as AltA. AltP events refer to introns overlapping with each other but with both 5'- and 3'-ends differing. For intron/exon comparisons, if an intron was completely covered by an exon, the event was classified as IntronR. If an exon was completely covered by an intron, the event was classified as ExonS. ExonS events involving terminal exons and the AltA/D/P events related to ExonS events were removed. The process and algorithm for identifying and analyzing AS events is described in more detail in 1 . AS events identified from cross-species EST alignment were labeled as "cross-species AS events". Correspondingly, the events from same-species EST alignment were referred to as "same-species AS events".

Conserved AS events were identified in two ways: (1) Comparing cross-species AS events with same-species AS events and other cross-species AS events from different species; (2) Identifying orthologous gene pairs between Mt and Lj and comparing their AS events. In the first method, the Mt genome coordinates of the AS events predicted from multiple species ESTs were compared. Only events with identical coordinates of an alternatively processed intron(s)/exon(s) were regarded as completely conserved. In the second method, the orthologous genes were identified by searching ESTs mapped in both Mt and Lj genomes. In some cases, orthologs in At and Os were identified by reciprocal BLAST using annotated protein sequences from At, Mt and Os. Gene structures and AS events of orthologous genes were then compared to identify conserved AS events.