Although the re-annotation process changed the majority of gene models, the number of protein-coding genes changed minimally, from 13,601 genes in Release 1 to 13,474 genes in Release 2 to 13,676 in Release 3, of which 13,379 are in the euchromatin (Table 1) and 297 in the heterochromatin [25]. However, the Release 3 gene structures have changed to contain more exons. The total number of unique exons in euchromatin, defined as having unique sequence coordinate termini, has increased 11% from 54,793 in Release 2 to 60,897 in Release 3 (see Table 1). The number of protein-coding exons has increased as well, from 50,667 to 54,934 (we define a protein-coding exon here as any exon containing CDS, even if the majority of the exon is UTR). The consequence is that the average number of exons per gene has increased from 4.1 in Release 2 to 4.6 in Release 3, which is very similar to C. elegans (4.5 [26]) and Arabidopsis (4.6 [27]) but significantly lower than H. sapiens (8.9, see, for example [28]).

A major contributor to the increase in exons is the increase in the number of protein-coding genes with identified 5' UTRs. One limitation of ab initio gene prediction programs is that they predict only open reading frames (ORFs): EST and full-length cDNA data are absolutely essential to identify UTRs. The expanded set of available EST/cDNA data led to a significant increase in the number of annotated genes and transcripts with 5' UTRs (from 57% of the genes in Release 2 to 76% in Release 3) and 3' UTRs (from 36% of the genes in Release 2 to 72% in Release 3; Table 1). These numbers reflect data availability: sequences from cDNA clones representing at least one transcript from approximately 78% of the genes in Drosophila were supplemented by a large number of additional 5' ESTs [29,30]. The length of the UTRs also increased with these new data (Table 1): the average 5' UTR length per transcript (for genes with annotated UTRs) increased by 30%, to 265 nucleotides, and the average 3' UTR length (for genes with UTRs) by 19%, to 442 nucleotides.

Four times as many genes in Release 3 (20%) as compared to Release 2 (5%) show alternative transcripts (Table 1). The vast majority of these are due to alternative splicing (an introduced bias; see Materials and methods), but 13% are due to alternative promoters and 6% to alternative polyadenylation sites. Alternative splicing results in the 26% increase in the number of protein-coding transcripts, and is largely responsible for a 14% increase in the number of unique protein species: from 13,922 in Release 2 to 15,848 in Release 3.

The changes in gene models also result in larger proteins. Proteins in Release 3 have a mean length of 552 amino acids and a median of 421 amino acids. This is an increase compared to Release 2, where the mean was 503 amino acids and the median 385 amino acids. The longest transcript and protein are encoded by dumpy (dp), which encodes a massive 69.7 kilobase (kb) mRNA and a 23,054 amino-acid polypeptide. The Dp protein is a component of the extracellular matrix, and appears to serve as an elastic adhesion molecule at cuticle-cell junctions, such as the epidermalcuticle interface [31].

The vast majority (94%) of the Release 3 annotations contain sequences that are present in exons from Release 2; however, only 63% of the unique peptides in Release 2 are unchanged in Release 3 (Table 1). Of the 15,848 unique Release 3 peptides, 8,769 (55%) are exact matches to Release 2 peptides. Reciprocally, 45% of the peptides are different from Release 2, emphasizing that, although the overall picture of the number and distribution of transcription units in the D. melanogaster genome remains largely the same, the new annotations include many changes to the protein products encoded by the genome.

The sequences of the vast majority of transposons in Releases 1 and 2 were composites derived from a number of copies of that transposon type. In Release 3, these composite sequences are replaced with the actual sequences present in the sequenced y¹; cn¹ bw¹ sp¹ strain for each individual element [17,18]. In all, 1,572 transposons are annotated in the euchromatic Release 3 genome (Table 2): 682 long terminal repeat (LTR) transposons, 486 LINE transposons, 372 terminal inverted repeat (TIR) transposons, and 32 foldback (FB) elements. These data include both full-length and partial elements. Details of these analyses are reported in an accompanying article [18].

Small, non-protein-coding RNAs are also included in this re-annotation. We searched for new tRNA genes using the program tRNAscan-SE [32] and Sim4 alignments to known tRNAs: 290 are annotated in the euchromatin (Table 2). Release 1 reported 292 tRNAs [3]; two tRNA genes were deleted as a result of sequence finishing resolving repeated regions of the genome.

Other non-protein-coding RNAs are limited, in general, to those already curated in the FlyBase database [1,2]. All 23 of the known microRNAs in Drosophila are located precisely in the Release 3 genome. We annotated the majority of the 45 small nuclear RNAs (snRNAs) involved in splicing, with the exception of the four snRNAs, K2a, K2b, K8, and K9 [33], for which there were no sequence or cytological data available. Of the 41 snRNAs supported by such data, we found that nine were redundant entries, and another five could not be identified at the previously specified cytological locations, possibly due to strain variation and/or inaccuracy in previous localization experiments. Thus, we precisely located by sequence alignment 28 snRNAs in the genome, including a new copy of the snRNA:U4atac gene in the 83A region.

All nine of the small nucleolar RNA (snoRNA) genes in FlyBase were identified by Sim4 alignment of sequence obtained from the literature. In addition, we incorporated data from Tycowski and Steitz [34] and located 19 more snoRNAs. Identification of other snoRNAs should be possible in the future with the use of algorithms like Snoscan, which looks for 2'-O-ribose methylation guide snoRNAs [35]; however, the program will have to be customized for Drosophila.

The longer non-protein-coding RNA genes αγ-element, bft, RNaseP:RNA, Hsr-omega, 7SLRNA, pgc, roX1, roX2, and iab-4, are annotated in the genome. 27 new 'miscellaneous non-coding RNA' genes were detected by alignment of spliced DGC cDNAs that did not appear to contain an ORF of significant length. In some cases these appear to be candidate antisense genes, which have also been reported in other organisms [36]. Further experiments will be necessary to verify the existence of these interesting genes and to determine their function.

The number of pseudogenes reported in Drosophila is substantially smaller than that in Caenorhabditis elegans [37,38]. We annotated the 12 pseudogenes in FlyBase that map to the euchromatic sequence and correspond to protein-coding paralogs (see Supplementary Table 1 in the additional data files). We identified five new pseudogenes: four histones and one lectin (CR31541) (Supplementary Table 1). Of these 17 pseudogenes, 15 are recombinationally derived (with introns, in tandem to their functional paralogs), one (Mgstl-Psi) is retrotransposed (with a poly(A) tail, lacking introns which its functional paralog possesses), and one is too degenerate to classify definitively. We did not make any attempt to comprehensively survey for new retrotransposed pseudogenes or annotate pseudogenes identified by Echols et al. [37]. WormBase [39,40] currently reports 392 pseudogenes in C. elegans. It is very likely that a subset of the genes identified as protein-coding genes in Release 3 are actually pseudogenes. In particular, 19 protein-coding genes were noted as containing a 'probable mutation in the sequenced strain' and more than 400 were marked 'problematic' because of inconsistencies with the experimental evidence and the predicted ORFs.

Release 3 contains a total of 802 new protein-coding genes (Table 1), that is, gene models that show no overlap with exons in Release 2. Of these, 55 (7%) are based solely on gene-prediction data, and 20 of these 55 are based on GENSCAN predictions alone. Unlike Releases 1 and 2, which relied heavily on Genie [3], Release 3 annotations did utilize GENSCAN predictions (with at least one exon with a score > 45) in the absence of other data. The majority of the new genes show matches to EST (573; 71%) or full-insert cDNA sequences (273; 34%), indicating the importance of these alignments in identifying new genes missed by the ab initio gene prediction programs. An additional set of new genes was identified by the community in error reports (52; 7%) or in GenBank/EMBL/DDBJ submissions (58; 7%). Finally, we created 338 (42%) new annotations in Release 3 using protein homology data from BLASTX analysis, arising from the comparison of translated Release 3 sequence with sequence of other proteins in Drosophila or other model organisms, in the absence of other supporting data.

Release 3 sequence 'finishing' had the largest impact on areas of repetitive sequence, because the Release 2 WGS sequence assembly often collapsed these regions [17]. Duplicated sequences present assembly challenges to genome sequencing efforts; tandemly duplicated genes tend to collapse in sequence assembly and cannot be annotated until the duplications are resolved. Whole-genome analysis of the Release 2 sequence suggested that Drosophila has fewer newly duplicated genes than nematodes or yeast [41]. We investigated whether sequence finishing might have uncovered previously undiscovered duplicated genes in Drosophila. From this analysis, we found that the number of newly duplicated genes is more similar to C. elegans and Saccharomyces cerevisiae than previously believed.

We measured the frequency of newly annotated duplicated genes by comparing each of the transcripts encoded by the 802 new genes in Release 3 to all Release 3 transcripts using the BLASTN program. Of the new genes, 124 (15%) have duplicate genes (75% identity, probability = 1 × e^-25) somewhere in the genome (whereas 10% of a random sample of Drosophila genes have duplicates by this measure). Thirty-six new genes are in repeat regions that were collapsed in Release 2 and have now been resolved. For example, in the previously annotated ten-gene trypsin cluster on chromosome arm 2R, three new trypsin genes (CG30025, CG30028, CG30031) have been added ([17] and Figure 1).

We rejected a total of 345 Release 2 genes during the Release 3 re-annotation (Table 1), primarily on the basis of a lack of supporting computational or experimental evidence (see Materials and methods). Nineteen Release 2 genes were deleted because they were contained within transposable elements. If based solely on computational gene-prediction evidence, genes that were less than an arbitrary length of 100 amino acids were deleted. We required an arbitrary length of 50 amino acids for all annotations not specifically supported by literature references (for example, the DIRG genes [42]), and 42 of the deleted Release 2 annotations were removed because they failed to meet this criterion (Figure 2a, inset).

The sizes of the predicted protein products in Release 2 and Release 3 were compared (Figure 2a), along with the protein sizes of Release 2 annotations deleted in Release 3, and sizes of proteins newly added in Release 3 (Figure 2b). When examining the size range of 0-50 amino acids, there is a marked decrease in Release 3 annotations compared to Release 2, due to the more stringent data requirements for small annotations in Release 3 (Figure 2b, inset).

Evidence supporting the merger of gene models came mainly from the alignment of full-length cDNA sequences and, to a lesser extent, from protein homology evidence. Merges based solely on BLASTX similarity were more difficult, as the exact exon-intron structure of the merged model was not experimentally indicated. A total of 1,351 Release 2 genes were merged to form 602 (5% of total) Release 3 genes. Sometimes the original predictions were spaced quite far apart in the genome, a probable reason that the gene prediction algorithm(s) separated the exons. For example, multiple ESTs support a merge of CG14409 and the Flotillin-2 gene (Flo-2 or CG11547), adding two 5' exons almost 80 kb away from the Release 2 annotation of the Flo-2 gene (Figure 3). The new Flo-2 transcript encodes a protein with an additional 50 amino acids at its amino terminus.

Gene model splits were often necessitated by the facts that gene-prediction programs such as Genie and GENSCAN can string together genes that lie close to each other and do not resolve nested genes. Of the Release 2 genes, 322 were split to form 675 (5% of total) Release 3 genes. For example, the annotated gene CG6645, with 5 exons in Release 2 (Figure 4), appears to have been based on a Genie prediction (note that GENSCAN had predicted two separate genes). EST evidence and BLASTX homology to other fly proteins indicated that this gene should be split into two three-exon genes, CG32054 and CG32053. One 5' UTR exon and one protein-coding exon in CG32053 were missed by both Genie and GENSCAN. Thus, neither Genie nor GENSCAN correctly predicted the structure of these two genes; each correctly predicted aspects of the gene models, but EST and BLASTX data were necessary to accurately determine the structure of the two genes.

Gene splits/merges were defined as changes involving more than one gene in both Release 2 and 3. While not common, splits/merges are interesting in that they involve simultaneous restructuring of multiple Release 2 annotations. One notable example is the split of CG8278 into the CG30350 and sns annotations (Figure 5). In this instance, BLASTX, GenBank/EMBL/DDBJ, and cDNA records indicate that the 3' half of CG8278 should be split off as a separate gene model (CG30350), while the GenBank:AF254867 record indicates that the 5' exon of CG8278 plus six other Release 2 annotations should be merged into the extensive sns annotation. There were 93 cases in which Release 2 annotations suffered a reassignment of exons more complex than a simple gene split or merge to generate Release 3 annotations.

The phenomenon of genes within genes, in which a gene is included within the intron of another gene, is common. In the analysis of the 2.9 Mb Adh region of Drosophila, the frequency of nested genes was reported to be approximately 7% [44]. In extending this analysis to the entire euchromatin, we find that 7.5% (1,038) of all Release 3 genes, including non-coding RNAs, are included within the introns of other genes. Of the 879 nested protein-coding genes, the majority (574) are transcribed from the opposite strand of the including gene. We observed 26 cases in which the exons and introns of a gene pair are interleaved. Transposons may also be located within the introns of genes; we observed 431 such cases.

We analyzed the mRNAs predicted for neighboring genes to find those transcripts that share common non-protein-coding genomic sequence. About 15% of annotated genes (2,054) involve the overlap of mRNAs on opposite strands. Some of these involve overlapping messages that have been previously described (for example, Dopa decarboxylase and CG10561 [45]); however, the vast majority were not previously known to overlap. Complementary sequences between distinct RNAs from overlapping genes on opposite strands have previously been reported in eukaryotes and have been implicated in regulating gene expression (for reviews see [46,47]). For example, the complementary sequence shared between Dopa decarboxylase and CG10561 is thought to be involved in regulating the levels of these transcripts [45]. The large number of such overlapping transcripts identified here raises the possibility that antisense interactions may not be an uncommon mechanism for regulating gene expression in Drosophila.

We were surprised to find over 60 cases of overlapping genes on the same strand, for which cDNA/EST data indicate that the 3' UTR of the upstream gene overlaps the 5' UTR of the downstream gene. In some instances, the 3' UTR of the upstream gene extends past the postulated translation start of the downstream gene. One example of such an overlapping model is CG9455 and Spn1 (CG9456), tandem genes encoding serine protease inhibitors (Figure 6). The two gene models are individually supported by a variety of BLASTX data as well as full-insert cDNA sequences. Interestingly, the 5' exon of the DGC cDNA clone covering the Spn1 gene (AT24862) is entirely included in the 3'-most exon of the CG9455 DGC cDNA clone (GH04125). The existence of overlapping genes raises many questions. Are such pairs of genes typically co-regulated? Where are the transcriptional regulatory elements for the downstream gene? What are the structural constraints on the overlapping sequences?

One mechanism for increasing potential protein and regulatory diversity is through the production of alternative transcripts. Approximately 20% of Release 3 genes have more than one predicted transcript, and this is almost certainly an underestimate. Many instances of internal alternative splicing as well as alternative polyadenylation will have been missed, as our dataset of cDNA sequences contained many more 5' ESTs than 3' ESTs or complete cDNAs. As cDNA collections are expanded, including those representing specific stages, tissues, and cell types, additional genes with multiple transcripts and additional protein species produced by alternative splicing will undoubtedly be identified. Despite likely underestimation, the level of alternative splicing that was observed clearly illustrates that alternative splicing is an important mechanism for generating transcript diversity in Drosophila (see Supplementary Table 2 in the additional data files).

Alternative splicing creates opportunities for diversity both at the level of gene regulation and of protein diversity. In Release 3, 35% of the 2,729 genes encoding multiple transcripts generate only one protein product; the transcripts differ only in their UTRs. Very commonly, these alternative transcripts vary in the location of 5' non-coding exons, suggesting the use of alternative promoters and offering the possibility of differential regulation. The other 65% of genes with alternative transcripts encode two or more protein products, indicating that alternative splicing generates considerable protein diversity in Drosophila.

A large number of related proteins can be produced from a single gene by the simple substitution of a single domain. This mechanism has been taken to an extreme level in the case of mod(mdg4), which produces at least 29 distinct transcripts that share 5' exons, but are alternatively spliced to an array of different 3' exons [48,49]. Remarkably, eight of these transcripts appear to be generated by a trans-splicing mechanism, using variable 3' exons encoded on the opposite strand. (Seven trans-spliced variants were previously reported [48,49]; our analysis suggests eight.) Although we did not find any further examples of trans-splicing, we did find that a similar gene, lola, generates at least 21 alternative transcripts (including four previously described [50]). The many lola transcripts also share 5' exons, but contain one of an array of different 3' exons. Both lola and mod(mdg4) encode families of specific RNA polymerase II transcription factors that include a BTB/POZ dimerization domain near each amino terminus [50,51]. mod(mdg4) has been implicated in a range of cellular and developmental processes, including chromatin insulator functions [52] and apoptosis [53], and it has been suggested that its many different isoforms underlie the pleiotropic nature of this gene [49].

Alternative splicing can produce two (or more) distinct non-overlapping protein products from a single pre-mRNA species; we identified 12 such cases (for example, Vanaso and α-Spec, see Figure 7). The mRNAs produced most commonly share 5' UTR sequences, but may also share 3' UTR sequences. FlyBase defines complexes of this type as two separate genes, since two non-overlapping protein products are produced. Although other groups sometimes describe such genes as dicistronic (since the unprocessed transcript is dicistronic), we do not include this type in our categorization of dicistronic genes (see below). The component coding regions are resolved on separate mRNAs, and thus internal translation initiation is not required. We view these cases as one extreme along a continuum of protein diversity created by alternative splicing.

Examples of dicistronic transcripts have been previously reported in Drosophila [54,55,56,57,58,59,60,61]. Our results confirm that, while not common, numerous examples of apparent dicistronic transcripts are encountered in Drosophila. We limit the term 'dicistronic' to genes that meet the following criteria: two distinct and non-overlapping coding regions contained on a single processed mRNA, requiring internal initiation of translation of the downstream CDS. In order to categorize a transcript as dicistronic, we required that each CDS exceed 50 amino acids in length and show some similarity to known proteins. The Release 3 annotation contains 31 gene pairs that can be described as dicistronic by these criteria (Figure 8, and see Supplementary Table 3 in additional data files). This includes 12 cases for which the dicistronic transcript is represented by a single cDNA. There are 17 additional pairs, denoted as putative, for which there is insufficient BLASTX evidence to support both ORFs in a dicistronic gene model (see Supplementary Table 3). Since the determination of genes as dicistronic requires multiple classes of data to confirm the transcript structure and validate the coding regions, there are undoubtedly additional dicistronic genes yet to be uncovered throughout the genome.

For many of the predicted dicistronic genes (31/48), there is evidence supporting alternative monocistronic transcript(s) for either the upstream or downstream CDS, or for both. This includes Mosc1A+Mosc1B, for which the monocistronic transcript encodes a fusion protein encompassing both CDSs [59]. In some cases the dicistronic form may be less prevalent than the monocistronic forms: it has been estimated that the dicistronic Adh+Adhr transcript is only 5% as abundant as that of the Adh monocistronic transcripts [57].

Translation of the second CDS of a dicistronic transcript requires that initiation of translation occur at an internal site. There are two proposed mechanisms for the initiation of internal translation. One mechanism is that internal initiation occurs by partial disassembly of the ribosome at the termination of translation of the first CDS, followed by continued scanning by the 40S ribosomal subunit [62]. The following conditions are thought to be criteria for the ribosomal scanning mechanism: an absence of any ATG codons in the intercistronic region, an intercistronic region of 15 to 78 bp, and an optimized consensus translation start site for the second CDS. We assessed the sizes of intercistronic regions and the number of ATGs in these regions, and found that there are seven pairs of dicistronic genes that appear to conform to this pattern (indicated in Supplementary Table 3). The majority of dicistronic cases clearly do not conform to such a model of partial ribosome disassembly and continued scanning. In four cases, the intercistronic region is less than 4 bp. However, most of the annotated dicistronic pairs are separated by several hundred base pairs, and the separation can be as much as 4.5 kb. In these longer intercistronic regions, there may be multiple ATG codons before the second translation start site. A mechanism of translation initiation utilizing an internal ribosome entry site (IRES [63]) appears a better explanation for these cases. Oh et al. [64] have hypothesized that certain Drosophila genes with long 5' UTRs might be translated via internal ribosome entry. If this is the case, translation of the second CDS within a dicistronic transcript may be effected by the same initiation mechanism.

Andrews et al. [16] suggested that the number of genes in Drosophila might be a severe underestimate, based on 7,297 testes EST sequences they generated and aligned to the annotated genome. However, using their data, as well as 23,087 additional testes-derived ESTs [29], we predict a similar number of genes in Release 3 as in previous releases. The more likely explanation for their results is that 5' exons, and not genes, were under-predicted in Drosophila, since there is EST evidence for testes-specific promoters and transcripts [29]. In most cases, the testes ESTs did not align to Release 1 genes because only the downstream CDS had previously been annotated for those genes, whereas in Release 3 the UTRs that match the testes ESTs are annotated.

Gopal et al. [15] reported 1,042 novel genes that were not included in the Release 2 annotation. After completing our re-annotation, we compared this set of 1,042 genes to our data. We found that 75% of their predicted genes mapped to euchromatin, 16% mapped to heterochromatin, 7% mapped within transposable elements, and 1% could not be found in the Release 3 genomic sequence. Of the 75% that mapped to euchromatin, 66% (520) do match Release 3 annotations. The remaining 34% did not match Release 3 annotations, leading to the possibility that some or all of these may represent novel genes. Incorporating their methods (threading GENSCAN predictions to look for structural homology) into our computational approach may uncover additional missed genes.

One way to address the quality of the Release 3 annotations is by comparative sequence analysis. In an accompanying paper, C. Bergman et al. [65] surveyed sequence conservation in approximately 0.5 Mb of the D. melanogaster genome containing 81 genes using comparative data from four Drosophila species (D. erecta, D. pseudoobscura, D. willistoni and D. littoralis). Their comparison to our D. melanogaster annotations detected no genes conserved in other species that were missed in Release 3 [65].

Other genes likely to be missed are genes with small ORFs (because of the arbitrary length cutoffs we used, see Materials and methods), and genes expressed transiently during development, at very low levels, and/or in cells and tissues not represented by the DGC cDNA libraries. Future DGC cDNA clones will be generated by directed screening of cDNA libraries with probes matching predicted exons [30], and cDNAs selected during the re-annotation process to represent alternative transcripts not currently in the DGC.

Of the final set of annotations, 93% contain sequences that are present in Genie-predicted exons and 96% contain sequences that are present in GENSCAN-predicted exons (Table 3). Only 249 (2%) of protein-coding gene models were created without an ab initio model, that is, solely on the basis of cDNA or protein homology evidence. The fact that 98% of our accepted annotations span a region containing a gene prediction supports both the strength of the prediction programs' algorithms as well as our reliance on them for our methods. However, both Genie and GENSCAN gene models were often wrong in detail, when compared to cDNA sequence alignments for three main reasons: first, exon mis-associations: the programs placed exons from one gene with the exons of a neighboring or nested gene; second, erroneous splice site calls: the donor and acceptor sites were slightly misplaced; or third, missed mini- and micro-exons: small ORFs were not identified [20,66].

The fraction of gene models that are based solely on gene prediction data has decreased considerably, from 2,348 (17%) in Release 1 to 815 (6%) in Release 3 (Table 3). This shift was primarily due to the more recently available Drosophila EST and cDNA sequences, rather than newly evident similarity to sequences in other species.

Alignment of full-length cDNA sequences from the same strain continues to be the best way to annotate gene models [66,67,68]. The number of ESTs generated by the BDGP project increased from around 86,000 in Release 2 to 246,248 in Release 3 [29] and the number of sequenced full-insert cDNAs from around 1,000 in Release 2 to over 9,000 in Release 3 [30]. For approximately 6,000 of these, the completely assembled sequence was available during the re-annotation effort (see Materials and methods). In addition, 8,699 ESTs from the community deposited in dbEST [69], including a set of 7,297 from a testes cDNA library [16], were available. In all, 78% of the protein-coding genes show a match to an EST sequence (Table 3) and over half to full-insert cDNA sequences. We anticipate further improvement to gene models as more cDNA data become available.

All introns within protein-coding genes were examined for conserved GT/AG splice junctions with the Sequin program [70,71], and all instances of annotations lacking GT/AG splice junctions were inspected and commented on. Of the 48,039 total splice junctions, 0.5% are annotated with GC/AG splice junctions, a frequency that might justify describing GC as an alternative splice donor. Eleven instances of AT/AC splice junctions are annotated. An especially well supported example of AT/AC usage is CG1354, which has more than 25 confirming ESTs. Other cases of non-consensus splice sites appear rare; however, more are likely to be documented in the future. The particular alignment algorithm used (see Materials and methods) and our reliance upon gene-prediction data imposed a bias against unconventional splice sites. In a number of cases for which an unconventional splice junction was supported by cDNA data, the precise location of the junction could not be determined, owing to repeated sequence at the donor and acceptor sites. A good example of this type of pattern is sba (CG13598). Two alternative transcripts for this gene are supported by cDNA data, and both appear to contain an unconventional, ambiguous splice junction. The two transcripts share the unconventional splice acceptor site; they differ in the location of the non-consensus splice donor site, but the two donor sites are identical in sequence.

We also examined every gene model with an intron less than 48 bp. The frequency of such introns in Drosophila is low; 32 are annotated in Release 3. There are several well supported examples of introns less than 45 bp, with at least two supporting cDNAs derived from the sequenced strain. These include mod(r) and csul, each with an intron of 44 bp, and CG11892, with a diminutive intron of 43 bp.

We used the SWISS-PROT and TrEMBL protein databases [13] and the PEP-QC software program [12] to validate the integrity of the annotations and to track consistency with previously published data (see Materials and methods). Of the 3,687 annotated peptides with a cognate in the curated SWISS-PROT/TrEMBL dataset, 75% (2,764) were of identical length and had more than 99% sequence identity. Curators examined each case with less than 100% sequence identity, and in some cases, annotation errors were detected and corrected. For example, translation start sites were shifted to the experimentally reported position, which in some cases was downstream of the predicted start. However, in most cases discrepancies appeared to be due to strain-specific polymorphisms or errors in the reported DNA sequence on which the SPTRreal entries were based (see Materials and methods). Given the high quality of the underlying Release 3 genomic sequence, we believe that in many cases the Release 3 annotation is more accurate than the sequences deposited by the community in SWISS-PROT and TrEMBL.

The amount of evidence attributable to each Release 3 gene model varies considerably, and therefore our confidence in these gene models, even the confidence in two alternative transcripts encoded by the same gene, may differ greatly. To estimate the reliability of a gene model, we developed a classification system that groups data into four categories: computational gene predictions; protein similarities; alignments of ESTs and other partial cDNA sequences; and alignments of full-insert cDNA sequences. One point was assigned to each data type that overlapped a given annotation, and a score of 1 to 4 was determined for each transcript, with 1 being the lowest and 4 being the highest confidence (see Materials and methods). As shown in Table 4, more than 80% of transcripts and more than 75% of genes were assigned a confidence value of 3 or 4. Thus, we have high confidence in a large proportion of the Release 3 gene models.

The Release 3 annotations should be much more consistent than in previous releases because fewer curators were involved, a defined set of rules was used, and additional validation steps were performed (see Materials and methods). We set a rigorous standard for the annotations by requiring attributable evidence for every gene model, for example, a gene prediction, an alignment to a GenBank/EMBL/DDBJ accession, or a curated personal communication to FlyBase. However, during the Release 1 analysis, Drosophila researchers annotated particular families of genes about which they were expert and for which they may have had specific unpublished information. Much of the evidence for these annotations was not released to the public domain and is not currently available, so some of the details of these gene models were lost in Release 3. The solution in these cases is for biologists in the community to continue to submit error reports to FlyBase to be curated by FlyBase as personal communications. The resulting set of annotations will be stronger because every gene has traceable evidence that is available in the database and is annotated according to a standard set of rules.

As is expected with such a complex analysis, rules cannot be expected to cover every eventuality. As a result, some of the annotations are based partially on curator judgment, introducing a potential source of inconsistency. Visual inspection and curator expertise were absolutely necessary in overcoming shortcomings of the automated processes such as identifying GC splice donors and sorting out complex gene models. It was also essential for annotating unusual cases, such as the dicistronic genes and overlapping gene models. Further, it should be noted that manual annotation is an iterative process. Subsequent to an initial annotation call, a set of automatic verification steps was carried out. Potential errors were reviewed and, where appropriate, annotations were modified as a result of the verification analysis.

The BDGP will continue to finish the remaining problematic regions of the euchromatic genomic sequence to high quality (see [17]), and focus efforts on refining the sequence of the heterochromatin [25]. Changes to the sequence will be submitted to GenBank/EMBL/DDBJ every 6 to 12 months.

Because sequence updates at the time of new releases will result in changes to the coordinate system for each chromosome arm and for GenBank/EMBL/DDBJ accession units, it will be particularly important for researchers to make note of specific release and version dates when providing sequence coordinates. FlyBase encourages researchers to refer to coordinates as associated with specific GenBank/EMBL/DDBJ accession and version numbers.

Future re-annotation will be on a gene-by-gene basis, rather than a survey of the entire genome. Future analyses will include new large-scale datasets, including the Anopheles gambiae genomic sequence, the D. pseudoobscura genomic sequence, and additional DGC cDNA sequences [30]. Changes to the gene models will occur more often than changes in the sequence, and will be reported in date-stamped updates of the GenBank/EMBL/DDBJ accessions and FlyBase records. Such changes are reflected in the feature annotations only and thus do not constitute new releases, as the underlying genomic sequence does not change.

FlyBase will also focus on the localization of many more annotation features to the genome view, such as regulatory elements, mutational lesions, rearrangement breakpoints, and P-element insertion sites. Many of these sequence features are already in the FlyBase genetic data tables and gene annotation reports, based on data from literature curation, computational analyses (for example [77]), and large-scale projects such as the BDGP Gene Disruption Project.

In the annotation of genes in Release 1 attributes of gene products were predicted with respect to their molecular functions, the roles they might play in biological processes, and their cellular locations, using the controlled vocabularies developed by the Gene Ontology (GO) Consortium [78]. These predictions were computational, using a program known as LOVEATFIRSTSITE written by M. Yandell [3]. Since then, FlyBase curators have assessed each of these annotations, retaining in FlyBase only those that were reasonably secure, and have re-annotated many genes with GO terms of higher granularity. This work, together with the curation of GO terms from the literature and sequence records, has resulted in 7,299 genes sharing 25,057 GO annotations. This analysis has not yet been repeated for the Release 3 gene products, but the curation of GO terms for all new genes and all split/merged genes is now in progress. When this annotation is completed we will have a benchmark for further automatic predictions of GO terms, using programs similar to LOVEATFIRSTSITE [3] and PANTHER [79].

The publicly available version of Genie, which does not utilize EST or BLAST evidence [20], predicted 13,794 genes on the finished sequence. GENSCAN predicted 19,189 genes. As reported previously [3,20], Genie appears to predict fewer false positives, perhaps because it has been trained on Drosophila sequences, whereas GENSCAN has only been trained on vertebrate datasets [21]. However, GENSCAN also shows greater sensitivity than Genie, identifying some real genes that Genie fails to find. To balance the false-positive and false-negative rates of GENSCAN, we used an empirical prediction score as a threshold, as done previously [44]. In the absence of other supporting evidence for a gene, we used GENSCAN predictions only when at least one exon had a score > 45; this is a stringent threshold, as 21% of the genes in Release 3 with full-length cDNA evidence do not contain any exons scoring > 45.

To detect proteins with significant sequence similarity, we used BLASTX to compare translated genomic sequence to peptides in other species included in SWALL [13,83], and TBLASTX to compare translated genomic sequence to virtual translations of the rodent UniGene set [84] and insect sequences in dbEST [69]. We also looked for sequence similarities to Drosophila peptides that had experimental verification, but not to those representing purely hypothetical or computational gene models (see below). Although the number of proteins in a public database like TrEMBL [13] has increased exponentially in the time between the Release 1 annotation in November 1999 and the Release 3 annotation in 2002 [85], the increased size of the protein datasets resulted in a 14% increase in the number of fly genes that produce proteins with similarity to other proteins. In March 2000, Adams et al. [3] reported that 9,612 (71%) of the 13,601 of the Release 1 genes showed a match to another protein. We now find that 10,996 (82%) of the Release 3 protein-coding genes show a match by BLASTX or TBLASTX (with expectation value less than or equal to 1 × e^-7). However, we note that the datasets we used were fixed before the release of the genomic sequence of A. gambiae [86], the only other dipteran (or arthropod) with a complete genome sequence. We expect that a higher percentage of Drosophila proteins will show sequence similarity to Anopheles proteins, because A. gambiae is more closely related to D. melanogaster than are the other available model organisms [86].

Prediction of gene models was made more rigorous by the increased availability of cDNA data. However, misleading alignments can be created by the presence of genomic DNA contaminants, cDNA clones containing two independent cDNAs co-ligated in the same plasmid vector (chimeras), and internal priming of cDNAs during library synthesis. cDNA clones derived from incompletely processed primary transcripts are not readily distinguishable from alternative splicing without experimental verification. Moreover, cDNA sequences designated as full length may actually be truncated; approximately 1,000 of the 9,000 full-insert sequences from the BDGP are probably not full-length [29,30]. The Sim4 alignment tool can make mistakes in determining splice site junctions or completely fail to align very small exons [24,67]; indeed, a small number of cases of failure to align microexons were identified by Stapleton et al. [30] when they compared the predicted translation products of cDNAs with those of Release 3 gene models. However, Haas et al. found est2genome and other alignment tools were, in general, not superior to Sim4 [24,67]. Despite these limitations, alignment of complete cDNA sequences is invaluable in detecting UTRs, alternative splicing events, detailed exon-intron structures, nested genes, and other key aspects of gene models.

Full-insert Drosophila melanogaster cDNA sequences came from a number of sources. The largest set of full-insert cDNA sequences came from the BDGP Drosophila Gene Collection (DGC) project [5,29,30]. Of the protein-coding genes, 9,297 (69%) show a match to full-insert sequences from the cDNA clones in the DGC, and in some cases, more than one DGC clone provided definitive gene models for alternatively spliced products. At the time of annotation, we had access to full-insert sequencing reads from 9,074 of the 10,910 cDNA clones, but only some 6,000 of these had been fully assembled. Gene models based on incompletely assembled cDNA clones were marked 'incomplete'. These gene models will be among the first annotations to be updated.

Sequences deposited in public databanks like GenBank/EMBL/DDBJ [6,7,8,9,10,11] by Drosophila researchers provided definitive evidence for a number of genes. For a subset of well-studied genes, FlyBase curators synthesized all of the available sequence and literature data into high quality Annotated Reference Gene Sequences (ARGS) that have been deposited in GenBank's RefSeq division [1]. These ARGS sequences correspond to 795 (6%) of the Release 3 annotations.

Other sequences came directly to FlyBase as error reports from the scientific community. FlyBase curated 636 reports with information about 1,094 genes as personal communications, and any sequences supplied in these reports were aligned to the genome. In all, 825 (6%) of the annotations overlapped these sequences (Table 3). Accurate annotation of three gene families in particular was greatly facilitated by sequence submitted in error reports: 85 cytochrome P450 monooxygenase genes (B. Dunkov, personal communication, FBrf0132129, FBrf0126925; D.R. Nelson, personal communication, FBrf0136021), 80 gustatory receptor genes (H. Robertson, personal communication, FBrf0141780; K. Scott and R. Axel, personal communication, FBrf0137428), and 61 odorant receptor genes (H. Robertson, personal communication FBrf0136024; C. Warr and L. Vosshall, personal communication, FBrf0128191).

The extent of each transcript and corresponding CDS was extracted from the 'Drosophila Genomic Sequence Annotations' file (in GFF format [87]), which is available [75]. The extent of overlap of each transcript against the supporting evidence used during the re-annotation was determined using an intersection algorithm to determine the annotations overlapped by particular types of evidence [12].

The evidence datasets used included: gene prediction based on Genie [20] and GENSCAN [21]; Sim4 alignments to EST and full-insert cDNA sequencing reads derived from the BDGP cDNA project [29,30], the earlier analysis of the Adh region [44], and dbEST (for example [16]); FlyBase ARGS [1]; GenBank/EMBL/DDBJ entries identified as Drosophila cDNA sequences [6,7,8,9,10,11] and error report submissions to FlyBase [1,2]; and BLASTX protein homology data. For a complete list of the evidence datasets and their description see [82]. Data were filtered using the Bioinformatics Output Parser (BOP), which also assembled all EST and full-insert cDNA sequence reads from a particular cDNA clone into a virtual assembly (BOP [12]). The Apollo tool displayed these assemblies with sequence gaps indicated differently from introns.

The algorithm used to assess relative annotation quality assigned one point for overlap of a gene prediction, either Genie or GENSCAN or both. One additional point was assigned for overlap with protein similarity data. The remaining datasets were considered in the following order and resulted in an additional one or two points: the cDNA and annotation data were analyzed to determine if any entry in this class spanned the entire length of the CDS; if so, an additional two points were assigned, and if not, GenBank/EMBL/DDBJ and error report entries were analyzed and if any spanned the length of CDS, two points were assigned; if none of these data classes corresponded to the full-length CDS, then the existence of partial cDNA data and/or overlapping EST data merited one point. Details of the rules for this classification system can be found at [82].

SWISS-PROT/TrEMBL validation of the translated models by PEP-QC is described below. Both annotated segments and chromosome arms were validated using the Sequin software tool from the NCBI [71], which found mistakes in exon-intron structure, start of translation, and ID duplication. We queried our dataset for proteins < 50 amino acids, CDS features making up less than 25% of the predicted transcript length, introns < 48 bp, and visually inspected each annotation in these classes making comments where appropriate. We checked annotations that overlapped transposable elements and tRNA genes, or appeared multiple times in the genome with duplicate identifiers. We verified that deleted Release 2 annotations had no independent evidence in literature-curated references. Finally, in order to allow the construction of a wild-type proteome from the mutant sequenced y¹; cn¹bw¹sp¹ strain, we replaced annotated sequences from known mutated genes (y, cn, bw, MstProx, LysC, Rh6) with RefSeq wild-type sequences from GenBank with an appropriate note.

All introns within protein-coding genes were examined for conserved GT/AG splice junctions with the Sequin program [70,71], and all instances of annotations lacking GT/AG splice junctions were inspected and commented upon. Splice junctions were based upon alignment of cDNA/EST sequence and, in the absence of such data, on gene prediction models. Even for transcript structures based upon EST data, the number of atypical splice junctions is probably an underestimate. The alignment algorithm used (Sim4) forced intron junctions to occur at GT/AG sites whenever possible, even at the expense of a several-base mismatch. This occasionally resulted in apparent early translation termination, in which case the annotator checked for a GC donor that would allow read-through. Other GC splice annotations were based on information in the literature or GenBank/EMBL/DDBJ records. With the exception of GT/AG junctions, we imposed a higher standard of verification for unconventional splice annotations: sequence data from a cDNA isolated from the sequenced strain, or multiple consistent ESTs.

The SWISS-PROT and TrEMBL protein databases [13] were used to validate the integrity of the annotations and to track consistency with previously published data. The SWISS-PROT Protein Knowledgebase [80] is a curated protein sequence database that provides a high level of annotation, a minimal level of redundancy and high level of integration with other databases. The TrEMBL database [81] contains the translations of all CDS present in the EMBL Nucleotide Sequence Database [8,9], which are not yet integrated into SWISS-PROT [80]. A non-redundant set of SWISS-PROT and TrEMBL Drosophila sequences was created, and sequences representing purely hypothetical or computational gene models (those corresponding to CG, BG, or EG genes in FlyBase) were excluded. The PEP-QC program [12] compared the resulting collection of 3687 D. melanogaster sequences (SPTRreal) to the annotated peptides using BLASTP [22]. Each gene was placed into one of four 'validation' categories: Perfect match to SPTRreal (annotated peptide of identical length with 100% sequence identity), Single AA substitutions (a