We obtained 173 million Solexa/Illumina RNA-Seq reads from mRNAs extracted from four tissues (leaf, root, stem, callus). Of these, 138 million reads could be mapped unambiguously with SOAP (Short Oligonucleotide Analysis Package) 8 to the Vitis vinifera genome sequence assembly 44. The mapped reads were contiged to build candidate exons, which we call 'covtigs' (for coverage contigs, that is, regions obtained by contiging adjacent positions with coverage depth greater than a threshold). Candidate junctions between covtigs were then tested using the unmapped reads. Finally, a graph approach was used to chain the exons through validated junctions into gene models (see Materials and methods; Figure 1). All possible chainings between exons were retained, which allowed the annotation of alternative splice forms. The covtigs that were not involved in any validated junction were discarded, implying that no mono-exonic transcripts were annotated. The procedure, which we named G-Mo.R-Se, produced 46,062 transcript models, clustered in 19,486 loci (an average of 2.4 transcripts per locus). A plausible coding sequence (CDS) was found for 28,399 models, clustered in 12,341 loci.

Covtig definition was essential for the subsequent testing of junctions to be efficient, especially with respect to the splits and fusions of exons (see Materials and methods). The splitting of exons into separate covtigs can occur when the read coverage depth goes down (below the depth threshold used for building covtigs), which can be due either to repeated regions (we only retained the reads that mapped at a unique position on the genome), to mismatches/gaps in the genomic sequence (we only kept the reads mapped with at most two mismatches and no indels), or to experimental biases leading to depth variations in the cDNAs sequenced and to non-normalization of the library. Indeed, some biases in the coverage uniformity of reads have been observed in previous RNA-Seq studies 23.

We aimed at correcting the splits in two ways. First, at the covtig definition step (step 1 in Figure 1), we extended the covtigs using all 16-mers found in the reads, in order to step over mismatches and short repeats. Then, at the model building step (step 4 in Figure 1), we fused together models that were linked by an open reading frame.

The artifactual fusing of exons into one single covtig can occur when the mRNA sample contains immature transcripts with retained introns, providing reads that map into the introns. Since the immature transcripts are expected to be under-represented in the set of mRNAs, the depth in the retained introns is expected to be lower than in the adjacent exons: setting an appropriate depth threshold for the building of covtigs should avoid such fusions.

The depth threshold used for covtig construction was set to balance the number of splits and the number of fusions. Indeed, low thresholds will generate few splits but numerous fusions, and conversely, high thresholds will generate few fusions but numerous splits. In order to correct more fusions, we could extend the testing of junctions inside the covtigs, instead of testing junctions only between covtigs.

We evaluated the direct mapping of reads, the initial candidate exons (covtigs), and the final models produced by G-Mo.R-Se at the nucleotide level in comparison to the reference V. vinifera annotation 44 (Table 1). The depth threshold set to build the covtigs discards most of the noise (63% of the nucleotides covered by reads are located in intergenic or intronic compartments compared to only 40% of the nucleotides covered by covtigs) while retaining the signal falling in exons (66% of the exonic nucleotides are covered by reads, and 56% are covered by covtigs). This noise is likely to correspond to transcriptional background, expression of transposable elements, or genomic contamination in the samples sequenced, rather than to SOAP mapping artifacts, since we only retained positions where reads could be mapped uniquely, with at most two mismatches. When considering final models instead of initial covtigs, the sensitivity decreases slightly (from 56% to 43% of exonic bases covered) but the specificity increases greatly (from 60% to 80% of the nucleotides - in covtigs or models - fall in the exonic compartment), suggesting that most of the covtigs that could not be linked to any other covtig resulted from transcriptional or experimental noise. The models still include about 1% of the nucleotides from the intergenic compartment, indicating that this compartment harbors new, previously unannotated, genes.

We managed to select a satisfying depth threshold with respect to the splits/fusions (Figure S1 in Additional data file 1), as well as the signal/noise ratios. Obviously, the optimal depth threshold will be highly dependent on the characteristics of the dataset analyzed, such as the complexity of the transcriptome, the amount of alternative splicing, the amount of transcription outside of protein-coding genes, and the sequencing depth, and must be carefully selected in order for G-Mo.R-Se to work optimally.

We compared the final G-Mo.R-Se models and the structures obtained by assembling the reads with Velvet 14 and mapping the assembled contigs to the genome with est2genome 37 (Table 2). Fewer reference genes are overlapped (on at least one nucleotide) by spliced Velvet contigs than by models (40.3% and 50.3%, respectively). The number of genes overlapped on at least 75% of their nucleotides drops even more for Velvet contigs compared to G-Mo.R-Se models (from 30.6% to 11.8%), indicating that most of the genes that are overlapped by Velvet contigs are not covered over their whole length. The average number of models or Velvet contigs per gene - 1.28 and 2.05, respectively - also reflects that the reference genes are more fragmented by Velvet contigs than by G-Mo.R-Se models. Additionally, we investigated the accuracy of the G-Mo.R-Se models and Velvet contigs on the structural point of view using a collection of cDNAs: 56% of the cDNA loci are predicted exactly (all exon/intron boundaries) by G-Mo.R-Se models, and 32% by Velvet contigs (Table S1 in Additional data file 1). We compared the average coverage depth of reference genes that are correctly annotated by G-Mo.R-Se models and Velvet contigs (that is, that have at least 75% of their nucleotides covered). A minimal depth of 4 is sufficient for G-Mo.R-Se models to annotate genes satisfactorily, whereas a minimal depth of 13 is required for Velvet contigs (Figure 2). Since G-Mo.R-Se relies on the genome sequence, no significant overlap between reads is necessary to put them together in a covtig: they just need to be adjacent on the genome. This explains why a much lower coverage depth is required for G-Mo.R-Se than for Velvet. Unlike direct assembly of reads, the G-Mo.R-Se pipeline is able to detect transcripts that are weakly represented in the reads set (either because they are weakly expressed or problematic to extract).

We compared the G-Mo.R-Se pipeline to a classic cDNA sequencing approach, using a reference set of 112,175 V. vinifera cDNA sequences from five tissues (including 87,199 multi-exonic cDNAs clustered in 7,895 loci) that were sequenced with the Sanger technology during the course of the V. vinifera genome sequencing and annotation project 44 (Table 3).

The 46,062 G-Mo.R-Se models overlap about 70% of the 7,895 cDNA loci (on more than 75% of their nucleotides). The most obvious reason why about 15% of the cDNA loci are not overlapped by any model is that they correspond to repetitive DNA. We compared the proportion of unique 32-mers (in the whole V. vinifera genome) for the 5,449 cDNA loci well covered by models and the 1,064 cDNA loci uncovered by models. It appears that most of the cDNA loci that were missed by models are mainly constituted of non-unique 32-mers (Figure 3). When considering only the 4,822 loci where all the 32-mers are unique, 95% of the cDNA loci are hit by a model (Table 3). Among the 5% of cDNA loci that are missed, some are too poorly covered by reads for covtigs to be built and/or junctions to be validated, and others have reads in their introns, which create fused exons, preventing the models from being detected as spliced, since one large covtig spans the whole locus.

Interestingly, G-Mo.R-Se detects 2.5 times as many loci as the standard cDNA sequencing approach (19,486 loci versus 7,895). Among the 19,486 G-Mo.R-Se loci, only 36% overlap cDNA loci. We compared the characteristics of the 5,698 G-Mo.R-Se loci that overlap cDNAs on at least 50% of their nucleotides and the 12,392 loci that are outside cDNA loci (Figure 4). The G-Mo.R-Se loci that are new with respect to standard cDNAs tend to be expressed at lower levels than the loci that overlap cDNAs. These loci are investigated in more detail in the section 'Identifying novel genes and improving gene annotation'. The RNA-Seq technology, combined with G-Mo.R-Se, is able to detect gene expression that would be scored silent with a standard cDNA cloning and sequencing approach, or would necessitate an extensive Sanger sequencing effort.

On average, we annotated 2.4 models per locus. By removing the redundancy (structures fully included in other structures; see Materials and methods) from the cDNA sequences, we retained 9,827 representative sequences, with an average of 1.25 transcripts per locus. The models appear to be capturing more alternative splice forms than the cDNAs. However, as we build all possible models that correspond to the longest possible paths going from one covtig to another through validated junctions, some of the models probably do not correspond to real transcripts (for instance, if they link alternative exons that are incompatible, like models M₃ and M₄ in Figure 1). Since the long-range splice contiguity can not be inferred from short reads, we quantified short-range alternative splicing events in the models (all models, and only CDS portions of coding models) and in the cDNAs 46 (Table 4).

The G-Mo.R-Se pipeline does not allow the detection of intron retentions (IRs), since we do not currently test junctions inside covtigs: if the depth in the retained intron is greater than the threshold we used to build the covtigs, we will get only one splice variant containing the retained intron. It is likely that most of the exon fusions we detected by comparison with the cDNAs (Figure S1 in Additional data file 1) correspond to cases of IRs. However, we were able to detect alternative donors or acceptors, skipped exons, and mutually exclusive exons. The relative abundance of these different classes of events is similar in the models and the cDNAs (from the most prevalent to the least prevalent: alternative acceptors/donors, skipped exons, mutually exclusive exons), but the total number of alternative splicing events in models (11,842 in all models, 5,152 in CDS portions) is much higher than in cDNAs (944 events, when removing the 1,227 IRs). The splice forms expressed at low levels, which could not be detected with cDNA cloning and Sanger sequencing, appear to harbor an unexpected number of alternative splicing events. It is likely that all these events are not compatible with the coding capacity of the transcripts. However, when restraining the analysis to the coding portions of models with plausible CDSs (that is, likely to be correctly predicted), the number of alternative splicing events remains higher than for cDNAs and the proportions of the different types of events remain unchanged. As an example, Figure 5 shows a locus where three alternative coding models were predicted: two of them (M₂ and M₃) are already supported by EST evidence, but the third model (M₁) corresponds to a novel alternative splice form. Although the number of alternative splicing events is higher in the RNA-Seq dataset than in the cDNA dataset, the proportion of loci where alternative splicing occurs is similar for cDNA clusters and G-Mo.R-Se models (10% and 8%, respectively). These results are in agreement with previous studies that showed that the fraction of alternatively spliced genes is lower in plants than in animals 47. Notably, of the 944 non-IR events detected in cDNAs, the models detect only 175 (18.5%): though some of these events might result from incorrect mapping of the cDNAs, most of them are likely to be real, and to have been missed by G-Mo.R-Se (Table 5). The pipeline detected only 7.2% of the skipped exons and 25% of the mutually exclusive exons, which is likely due to the limited number of neighboring covtigs (20) we tested to validate the junctions. Only 22.6% of the alternative donors/acceptors were detected because we searched for junctions only 100 nucleotides around the covtig boundaries, which limited the window where alternative splice sites could be discovered (see Materials and methods). Obviously, the model construction was not designed to capture the whole alternative splicing landscape of a genome. But still, the non-exhaustive view that we obtain is much richer than what could have been suspected from classic EST sequencing. In order to study alternative splicing exhaustively, which is out of the scope of this study, specific tools will need to be developed.

Expectedly, since V. vinifera belongs to a phylogenetic branch where a profusion of resources are available, most of the models (95%) that fall outside of cDNAs overlap the reference annotation 44, or other resources such as GeneWise hits with Uniprot proteins 3949, and ESTs from other species (Table S2 in Additional data file 1). However, 675 models are completely novel, or 116 when considering only models with a plausible CDS. We compared the characteristics of the models that are novel and the models that are supported by evidence, which we now call 'known' models (Table 5).

The proportion of models with a plausible CDS drops when considering the novel models compared to the known models (from 65% to 17%), as well as the average number of exons per model (from 8.2 to 2.3 for all models). It is likely that some of the novel models correspond to false predictions: if one junction is validated erroneously, it will create a false two-exon model. Nevertheless, the proportion of models with more than two exons is higher in the subset of novel models having plausible CDSs (53%) compared to all novel models (17%), which suggests that at least some of them are genuine novel coding loci. In addition, the novel loci that are non-coding could either correspond to coding transcripts that were mis-annotated by the pipeline (wrong splice site generating a frameshift, models associating incompatible exons), to coding transcripts where no CDS could be detected because of frameshifts in the genomic sequence, to genuine non-coding transcripts, or to transcriptional/experimental noise (Figure 1). The structure of one of the novel models, spanning eight exons, is shown in Figure S2 in Additional data file 1. A Blast 48 search against Uniprot 49 revealed an homology to a transcription regulator from Arabidopsis thaliana. The homology was below the sensitivity threshold required to map proteins to the genome during the annotation process. In addition to the discovery of novel splice forms and novel loci, G-Mo.R-Se models enrich the reference annotation by extending (in 5' or 3') about 40% of the reference genes they hit. G-Mo.R-Se models thus constitute a valuable resource for improving V. vinifera gene annotation.

RNA-Seq reads were obtained (as described in Del Fabbro et al., unpublished data) by sequencing cDNA obtained from four tissue samples with the Solexa/Illumina technology: leaf (11 lanes), root (9 lanes), callus (9 lanes), and stem (9 lanes). The mRNA molecules were purified from total RNA extractions and fragmented before cDNA synthesis (with random hexamer primers). The protocol was not strand-specific. The single-end reads obtained were 32 nucleotides long, except for 5 lanes in the callus sample, where the reads were 35 nucleotides long. The resulting 172,545,778 usable reads (5.4 Gbases) were mapped to the V. vinifera genome 44 using SOAP 8 with a seed length of 12 and default parameters: 138,326,238 reads (4.6 Gbases) were mapped at one unique position with at most two mismatches and no indels. As a consequence, reads that align to exon-exon junctions could not be mapped to the genomic sequence.

The G-Mo.R-Se method for building gene models from short reads is summarized in Figure 1. The first step is the definition of covtigs (coverage contigs). They are built by contiging the positions where short reads are aligned above a certain coverage depth threshold. This threshold is a parameter that needs to be adjusted in order to balance sensitivity and specificity as well as splits and fusions. In the absence of a training set to quantify the splits and fusions, this parameter can also be optimized by maximizing the number of junctions validated in the next step. Before the subsequent testing of junctions, the covtigs were extended using all 16-mers found in short reads, in order to step over mismatches and short repeats. It is important to note that the read length limits the detection of very short exons (< 35 nucleotides).

In the next step, we searched for donor (GT or GC on the forward strand, and AG or AC on the reverse strand) and acceptor (AG on forward strand and CT on reverse strand) splice sites 100 nucleotides inside and outside each covtig boundary. This enabled us to create a list of oriented candidate exons (with putative alternative donor and/or acceptor splice sites) for each covtig.

The third step was the validation of junctions between candidate exons using unmapped reads, since reads that align to exon-exon junctions were not mapped to the genomic sequence. We tested all candidate exons derived from a given covtig with the candidate exons derived from the 20 next covtigs. All the putative junctions were tested using a word dictionary approach. The dictionary (with a word size of 25) was built using the unmapped reads. Ten words (8 nucleotides on the first exon and 17 nucleotides on the second exon, 9/16, 10/15, 11/14, 12/13, 13/12, 14/11, 15/10, 16/9, 17/8) were derived from each putative junction, and their presence in the dictionary was tested. In order to validate a junction, at least five different words need to be found in the dictionary, and the total number of occurrences of all words derived from each junction needs to be of the same order of magnitude as the average depth of the adjacent covtigs (greater than 1/10 of their average depth).

The efficiency of the junction validation procedure relies on the covtig definition step for the following reasons: only the junctions between each covtig and the 20 next covtigs are tested, meaning that if more than 20 'false' covtigs are defined between 2 'real' covtigs, the junction between the two real covtigs will not be tested; only 100 nucleotides around the covtig boundary are scanned for putative splice sites, meaning that if the covtigs are too short or too long, the correct junction will not be tested; only junctions between covtigs are tested, meaning that if a covtig corresponds to a fusion between two exons, the correct junction will not be tested, and the final model will include a retained intron. On the other hand, if an exon is split between two covtigs, no junction will be valid between those covtigs, leading to the splitting of a gene into separate models. As a consequence, in the absence of a training set (annotated genes, ESTs, and so on) to calibrate the depth threshold used for building covtigs, it is possible to optimize the threshold by maximizing the number of validated junctions. G-Mo.R-Se can thus be used for de novo annotation.

For the last step, the model construction relies on the graph of candidate exons linked by validated junctions on the same strand. The models correspond to all the longest paths linking candidate exons through validated junctions. Candidate exons that are not involved in any validated junction are discarded, implying that no mono-exonic models are produced. In order to correct potential gene splits, we fuse together adjacent models (on the same strand) that are linked by an open reading frame.

Additionally, all models produced by G-Mo.R-Se are searched for CDSs. When the longest CDS (if greater than 50 amino acids) spans at least two-thirds of the nucleotides of a model or the number of non-coding exons is lower than the number of coding exons, the CDS is qualified as plausible. Models with plausible CDSs are likely to correspond to protein coding genes. Plausible CDSs could be detected for about two-thirds of the models. The G-Mo.R-Se models can be downloaded from the G-Mo.R-Se website 45 and visualized on the V. vinifera genome browser 50.

The same clustering procedure was applied to models and cDNA sequences aligned on the genome. We used a single linkage clustering approach, where a link between two models was created if they had a cumulated exonic overlap (on the same strand) of at least 100 nucleotides (only overlaps of at least 10 nucleotides were considered). A graph-based approach was used to resolve the single linkage clustering. Additionally, the redundancy was removed from the cDNAs by discarding all transcript structures that were fully included in longer structures. We detected all pairwise alternative splicing events between intron pairs, with the same method as described in 46. All tandemly duplicated genes were discarded from the alternative splicing events detected, since such genes may be artificially linked by cDNA mapping as well as model construction, and would generate false alternative splice forms spanning several loci instead of one. However, it is notable that, since the pipeline builds all possible models, it will always predict the two separate correct models in addition to the incorrect joined model(s).