On the basis of genome size considerations and the desire to investigate a range of divergence times in the genus Drosophila, we constructed fosmid libraries (approximately 40-kb inserts) for D. erecta, D. pseudoobscura, D. willistoni and D. littoralis (Figure 1). D. littoralis is closely related to the well-studied species, D. virilis, but has been reported to have less dispersed repetitive DNA than D. virilis (Kevin White, personal communication). We designed degenerate PCR primers for a set of eight well-characterized genes (apterous (ap), even-skipped (eve), fushi-tarazu (ftz), twist (twi), and Rhodopsins 1, 2, 3 and 4 (Rh1, Rh2, Rh3 and Rh4)) to obtain species-specific sequence-tagged sites (STSs) that were subsequently used for hybridization to gridded fosmid filters (see Materials and methods). Positive clones from the library screen were verified by PCR and restriction mapped to choose the longest clone containing the candidate gene and its regulatory regions.

In the initial design of this project, comparative sequence data was to be collected from a D. virilis P1 library [15]. Using a PCR-based plate-pool screening strategy, we isolated a P1 clone from this library containing an 83.2-kb insert from the ap region of D. virilis. Sequencing of this clone revealed long stretches of repetitive DNA, which complicated both assembly and comparative analyses. In addition, the insert size of the D. virilis P1 library (approximately 60-80 kb) was greater than necessary for comparative analysis of single gene regions. This clone was used to guide transgenic reporter analysis (see below), but has not been included in the other analyses reported here.

In total, 30 fosmid clones were isolated and sequenced using methods described in [9] which sum to 1,257,069 bp. All clones were finished to an estimated error rate of fewer than 0.17 errors per 10 kb, with an average estimated error rate of 0.03 errors per 10 kb. The lengths of fosmids sequenced for the eight candidate regions are shown in Table 1. Though we were able to obtain species-specific STSs for the D. willistoni twi gene, we were not able to obtain clones for this region from the D. willistoni fosmid library. We were also not able to obtain a species-specific probe for D. willistoni ftz, nor could we obtain any D. willistoni ftz clones using probes from other non-melanogaster species. Also shown in Table 1 are the lengths and locations of D. melanogaster genomic regions corresponding to the union of the Release 3 sequences homologous to all four non-melanogaster species. The union of sequences from all non-melanogaster species for the eight candidate regions covers 494.6 kb of the D. melanogaster genome; an additional 65.3 kb of D. melanogaster genomic sequence was sampled owing to rearrangements in non-melanogaster species. Thus the 1.25 Mb of comparative data presented here span over 0.5 Mb of coding and non-coding sequences of the D. melanogaster genome.

The 30 non-melanogaster fosmid (and the D. virilis ap P1) sequences were computationally processed using the pipeline used to re-annotate the D. melanogaster genome [12]. The only major modification to this pipeline was to add an additional tier of evidence containing the results of TBLASTN searches of all Release 3 D. melanogaster peptides [14] against non-melanogaster sequences. Predicted coding sequences were manually verified and refined using the Apollo annotation tool [13]. As no expressed sequence tag (EST) information exists to annotate transcribed non-coding sequences (such as UTRs) for the four non-melanogaster species, we annotated only protein-coding gene and exon models. Thus, in keeping with other gene-prediction studies (for example [16]), we use the terms gene and exon to refer to the translated components of genes and exons.

In the 30 fosmids, we predict a total of 164 protein-coding genes in non-melanogaster species (53 in D. erecta, 41 in D. pseudoobscura, 39 in D. willistoni, 31 in D. littoralis) that form orthologous clusters with 81 D. melanogaster genes. Of the 81 genes, 30 are 'known' genes that have been functionally characterized in some way by the community of Drosophila researchers; the remaining 51 genes are 'predicted' genes based only on the evidence in the Release 3 annotations ([14] and see Supplementary Table 1 in the Additional data files section). Of the 164 genes predicted in non-melanogaster species, 133 (81%) are full length; the remaining 31 (19%) are partial coding sequences that span the edge of the sequenced genomic clone. In non-melanogaster species, we predict 495 coding exons (148 in D. erecta, 133 in D. pseudoobscura, 111 in D. willistoni, 103 in D. littoralis) that form orthologous clusters with 264 D. melanogaster coding exons. On average, there are approximately two non-melanogaster species sampled per orthologous gene and coding exon cluster. Fifteen genes (10 complete) and 39 coding exons were sequenced in all four non-melanogaster species.

Qualitatively, our data reveal that the majority of D. melanogaster Release 3 gene models are highly conserved in divergent Drosophila species. This made it possible to automatically identify orthologous genes in non-melanogaster species using TBLASTN results in conjunction with Genie [17] and/or GENSCAN [18] predictions to improve intron-exon boundaries and identify small/divergent exons in Apollo. In the few discrepant cases where no clear ortholog could be unambiguously identified (such as the four closely related members of the Rhodopsin gene family), we used the conserved microsyntenic gene orders maintained in these species to resolve orthologs (see below). With the exception of the retrotransposition events discussed below, the intron-exon structure of gene models is highly conserved as well: only one case of intron gain was observed in the D. littoralis Rh2, as has been reported previously for Rh2 in the closely related species, D. virilis [19]. For a small class of genes (BcDNA:LD21213, Gr59a, Gr59b, CG9895, CG10887, CG17186, CG4733), orthologs could be identified in divergent species, but amino-acid sequences could not be reliably aligned with the D. melanogaster gene model. In addition, orthologs of four genes (CG13029, CG14294, CG12133, CG12378) could not be identified in non-melanogaster species except in D. erecta, the species most closely related to D. melanogaster. The absence of these genes is not simply due to insufficient sampling, since in these cases both 5' and 3' neighboring genes could be identified in more divergent species (see Figure 2, for example). These may represent genes overpredicted in both D. erecta and D. melanogaster, lineage-specific genes which evolved before the divergence of D. melanogaster from D. erecta, or genes which have transposed from (or to) other locations in the genomes of the more divergent species.

We used an evolutionary genetic approach, the K_a/K_s test, to assess the accuracy of these gene and exon predictions [20]. This test relies on the assumption that functionally constrained protein-coding sequences should exhibit significantly lower rates of evolution in amino-acid-encoding nucleotide sites (typically first and second positions in a codon) relative to silent sites (typically third positions in a codon). Quantitatively, this leads to the prediction that the ratio of the average rate of amino-acid substitution per site (K_a) relative to the average rate of silent substitution per site (K_s) for functionally constrained coding sequences should be significantly less than 1 [21]. Genes or exons which have a K_a/K_s ≈ 1 are inferred to evolve in the absence of functional constraint; genes or exons which have a K_a/K_s > 1 are inferred to evolve under the influence of positive selection. The significance of a K_a/K_s ratio can be determined by a likelihood ratio test of the probabilities of the data under the alternative hypotheses of functional constraint relative to no constraint [22]. Genes or coding exons with a K_a/K_s ratio significantly less than 1 'pass' the K_a/K_s test; genes or coding exons with a K_a/K_s ratio not significantly less than 1 'fail' the K_a/K_s test. The power of this test to detect functional constraint is influenced both by evolutionary distance and sequence length [20]; thus we analyzed both genes and coding exons in pairwise comparison with all four non-melanogaster species.

All pairwise gene-level comparisons studied here exhibited K_a/K_s ratios less than one (see Supplementary Table 1 in the Additional data files section). One hundred and fifty-five of 164 (94.5%) of these K_a/K_s ratios were significantly less than 1, indicating that the vast majority of genes in our sample show evidence of functional constraint. All nine pairwise comparisons that fail the K_a/K_s test at the gene level were D. melanogaster-D. erecta comparisons, and eight out of nine involved predicted genes (Supplementary Table 1). Genomic sequences for six of the nine genes which fail the K_a/K_s test at the gene level were sampled in more divergent species: four of these six genes could be identified in more divergent species (Lmpt, CG10887, CG14292, and CG4468), whereas two could not (CG12378 and CG14294), indicating that genes conserved in divergent species can fail gene-level K_a/K_s tests in comparisons among closely related species like D. erecta. Of the four genes identified only in D. melanogaster and D. erecta and not in more distantly related species, two pass CG12133 and CG13029) and two fail (CG12378 and CG14294) the gene-level K_a/K_s test. We note that of these four genes, the two genes that pass (CG12133 and CG13029) have multiple exons, whereas the two genes that fail (CG12378 and CG14294) have only a single exon. This result indicates that at least some of the genes found only in D. melanogaster and D. erecta are likely to be real genes under functional constraint.

Though the majority of pairwise exon level comparisons have K_a/K_s ratios less than one (Figure 3), a much lower proportion of pairwise comparisons at the exon level pass the K_a/K_s test. In total, 71.9% (356/495) of pairwise comparisons at the exon level pass the K_a/K_s test: 54.0% (80/148) for D. erecta; 78.9% (105/133) for D. pseudoobscura; 81.1% (90/111) for D. willistoni; and 79.6% (82/103) for D. littoralis. Coding exons from known and predicted genes pass the K_a/K_s test at similar rates: overall, (72.2% known versus 71.1% predicted), D. erecta (56.6% known versus 50.0% predicted), D. pseudoobscura (80.4% known versus 75.6% predicted), D. willistoni (81.5% known versus 82.1% predicted), D. littoralis (78.0% known versus 80.6% predicted). The majority of exons that fail the K_a/K_s test still have K_a/K_s ratios less than 1; only six non-significant pairwise exon comparisons (one in D. erecta, one in D. pseudoobscura, two in D. willistoni, and two in D. littoralis) have K_a/K_s ratios greater than 1 (Figure 3). As with gene-level comparisons, the most closely related species, D. erecta, fails the highest proportion of exon-level K_a/K_s tests. In contrast to gene-level comparisons, there is no tendency for exons from predicted genes to fail K_a/K_s tests relative to exons from known genes.

Pairwise comparisons that do not pass the K_a/K_s test could result from misannotated exons or an insufficient amount of divergence to resolve differential rates of amino acid and silent site evolution. Failure to pass exon-level K_a/K_s tests because of insufficient divergence is a function of divergence time and exon length [20]. Our results suggest that both factors contribute to non-significant exon-level K_a/K_s tests between species in the genus Drosophila. The fact that the most closely related species, D. erecta, fails the highest proportion of gene-and exon-level K_a/K_s tests indicates that insufficient divergence time contributes to non-significant comparisons. Exon length is also a factor, as there is a tendency for shorter exons to fail K_a/K_s tests in our data. For example, the average length of all exons failing the K_a/K_s test in comparisons between D. melanogaster and D. pseudoobscura is 22.1 codons, and the average length of all exons passing the K_a/K_s test is 152.1 codons. Similar results are obtained for pairwise comparisons involving D. melanogaster and D. erecta, D. willistoni or D. littoralis, and for both known and predicted genes (data not shown).

To determine if insufficient divergence time is the major cause of non-significant exon-level K_a/K_s tests, we performed multi-species exon-level K_a/K_s tests that capitalize on a greater total amount of divergence for a given exon [20]. The question addressed by this analysis is: does the addition of a third species to D. melanogaster-D. pseudoobscura pairwise comparisons increase the proportion of exons that pass exon-level K_a/K_s tests? For this analysis, we analyzed exons that failed pairwise tests between D. melanogaster and D. pseudoobscura using triplets involving D. melanogaster, D. pseudoobscura and one other non-melanogaster species. Using the same cutoffs for the pairwise exon-level analyses and a guide tree based on Figure 1, we tested 16, 14 and 13 exons which did not show evidence of functional constraint between D. melanogaster and D. pseudoobscura, for which we have sequence data available in D. erecta, D. willistoni and D. littoralis, respectively. Only 2 of the 16 (12.5%) non-significant exon-level D. melanogaster-D. pseudoobscura comparisons pass the K_a/K_s test when D. erecta is included as a third species, whereas 6 of 14 (42.8%) and 6 of 13 (46.1%) pass when D. willistoni and D. littoralis are included as a third species, respectively. These results demonstrate that multiple comparisons among divergent species can reveal functional constraint acting on coding exons that cannot otherwise be detected in pairwise comparisons.

Finally, as a preliminary assessment of the relative utility of A. gambiae genome sequences for comparative gene prediction in Drosophila, we attempted to identify homologs in A. gambiae of the 81 genes for which we have comparative sequence data in Drosophila. For 21 of the 81 genes in our study (25.9%) we were not able to obtain a clear homolog (defined as a high-scoring pair (HSP) with an expected (E) value less than 10^-20 and greater than or equal to 30% identity over 100 amino acids using default parameters of TBLASTN) in the A. gambiae mapped scaffold sequences; 11 of these 21 genes did not yield any HSPs at all. These results are compatible with a recent whole-genome analysis showing that 18.6% of D. melanogaster genes have no clear homolog in A. gambiae [23]. No clear homolog could be identified in the A. gambiae genome sequences for three of 30 (10.0%) known genes in our dataset, whereas a greater than three times higher proportion of predicted genes, 18 of 51 (35.3%), had no clear homolog in A. gambiae. Five D. melanogaster genes - the four members of the Rhodopsin gene family and CG5245 - have multiple HSPs in the A. gambiae genome sequences. We were able to resolve orthology for Rh4 only, as the sina gene in A. gambiae is contained within the Rh4 gene as in D. melanogaster and other species in the subgenus Sophophora.

Using the gene predictions discussed above as orthologous markers, we addressed the question of whether the microsyntenic relationships in the D. melanogaster genomic sequence surveyed are conserved in non-melanogaster species. In general, our data indicate that the microsyntenic order of coding and non-coding sequences is highly conserved in the genus Drosophila at the scale of individual fosmids (approximately 40 kb). Our data provide evidence for only six genomic rearrangements in these sequences occurring in the phylogeny of these five species, one each in the lineages leading to the D. littoralis ftz, D. pseudoobscura Rh1, D. willistoni eve, Rh1, and Rh3 regions, as well as in the ancestor of the D. melanogaster/D. erecta eve region (see Figure 1). All of these unique events occurred in non-coding intergenic regions and none of the rearrangement breakpoints is associated with detectable transposable element sequences (see also [24]). Although it is difficult to estimate the length distribution of microsyntenic regions in Drosophila from our data, it is clear that very small microsyntenic regions can be delimited in the Drosophila genome through multiple species comparisons. For example, the two independent rearrangements in the vicinity of the eve locus reduce this microsyntenic region to a approximately 20-kb interval of the D. melanogaster genome containing only three neighboring genes (Adam, CG12134 and eve) and their flanking non-coding sequences (Figure 2).

We can directly confirm the nature of one rearrangement (D. littoralis ftz) as a paracentric micro-inversion since both breakpoints are contained within a single fosmid clone. In this case, a small (approximately 14 kb) region containing the ftz coding sequence and flanking non-coding DNA is inverted between the Antp and Scr genes relative to D. melanogaster. Maier et al. [25] provide hybrization data for a similar rearrangement in the ftz locus of D. hydei, another member of the subgenus Drosophila. It is likely that the other rearrangement breakpoints we observe also result from paracentric inversions, the predominant form of genome rearrangement in Drosophila [26]. Consistent with this is the fact that rearranged sequences can be inferred to come from the same chromosome arm. At least two other breakpoints (in the D. willistoni Rh1 and Rh3 regions) also have probably arisen from micro-inversions, as in both cases only two genes are inferred to have switched order locally on the chromosome.

We also identified eight examples of novel genetic elements in non-melanogaster species, seven of which occur in intergenic regions (Figure 1). Four of these cases involve the insertion of novel transposable element sequences: full length Bari-1-like elements in both the D. pseudoobscura Rh1 region and the D. willistoni Rh3 regions, a partial I-like element in the D. willistoni Rh4 region, and a partial blastopia-like element in the D. littoralis Rh3 region. Identification of Bari-1-like transposon sequences in D. pseudoobscura and D. willistoni is consistent with previous observations [27]; I-like elements have been shown to exist in the melanogaster and obscura species [28], but this is the first report of I-like elements in the willistoni group. The other four cases arise from gene transposition including: a homolog of the D. melanogaster X-chromosome gene CG2222 in the D. willistoni eve region; a homolog of the D. melanogaster 3R-chromosome gene CG5245 in the D. pseudoobscura Rh1 region; a homolog of the D. melanogaster 3R-chromosome gene CG8319 in the D. willistoni Rh1 region, and the Rh4 gene in D. littoralis (see below). The CG5245-like gene in D. pseudoobscura and the CG8319-like gene in D. willistoni both are located in the same intergenic region between the Arc42 and PK92B genes, but result from independent events since they involve different ancestral sequences and occur on opposite strands in this intergenic region. This result suggests the possibility of hotspots for gene transposition in the Drosophila genome.

At least one novel gene, the CG2222-like gene D. willistoni, is likely to have originated from a retrotransposition event as this gene lacks introns while its closest homolog, found on a different chromosome arm in the D. melanogaster genome, has two introns. Another striking example of retrotransposition involves the D. littoralis Rh4 gene and illustrates the fact that functionally important genes can undergo dramatic changes in location and gene structure during genome evolution [29]. This gene maintains its microsyntenic relationship with neighboring genes in the 72D2-3 region of the D. melanogaster genome in Sophophoran species, but has retrotransposed into the intron of another gene, CG10967, in a region of the D. littoralis genome that corresponds to the 69E1-2 region of the D. melanogaster genome. As a result, genes contained in the intron of the Sophophoran Rh4 (sina, CG13030 and CG13029) have been lost in the process. Cytological evidence for transposition of Rh4 exists for the closely related species D. virilis and the more distantly related species D. repleta [29,30].

In contrast to the stability of microsyntenic gene order in the genus Drosophila, we found that the sample of genes studied here are scattered widely throughout the Anopheles genome. For example, of the 55 Drosophila genes that had a single clear homolog in Anopheles, 27 are located on D. melanogaster chromosome arm 2R. Of these 27 genes, ten, five, six and six are located on A. gambiae chromosome arms 2L, 2R, 3L and 3R. These results are consistent with previous reports comparing the locations of genes in D. melanogaster with A. gambiae, which indicate that extensive genome rearrangement has occurred since the divergence of these two lineages [23,31,32]. Some D. melanogaster genes in our sample do maintain microsyntenic relationships in A. gambiae, such as the Rh4 and sina genes. In this case, conservation of microsynteny is most probably maintained because of the nested relationship of these genes, and this configuration in the outgroup Anopheles supports the scenario that transposition of Rh4 occurred at some point in the lineage leading to the Drosophila subgenus (see above, and [29,30]).

In addition to providing a useful resource for studying comparative gene prediction and genome rearrangements, our data confirm and extend emerging trends in Drosophila coding sequence evolution. Table 2 summarizes the average rates of amino-acid and silent site substitution for all, known and predicted genes in our dataset. These data show that predicted genes tend to have a higher rate of amino-acid substitution than known genes in the genus Drosophila. This trend is significant for the three most closely related pairwise comparisons (D. melanogaster versus D. erecta, D. pseudoobscura or D. willistoni) but non-significant in the comparison involving the most distantly related species (D. melanogaster versus D. littoralis). No significant differences were detected in the rates of silent site substitution between known and predicted genes in any pairwise comparison, although predicted genes in D. pseudoobscura, D. willistoni and D. littoralis tend to show elevated rates of silent site substitution.

In contrast to expectation, average rates of amino-acid substitution are highest in comparisons between D. melanogaster and D. willistoni, not D. melanogaster and D. littoralis (Table 2, Figure 1). The overall increased rate of amino-acid substitution for genes in the D. willistoni lineage is caused by an increased rate of amino-acid substitution in predicted genes. For known genes, average rates of amino-acid substitution are consistent with the accepted phylogenetic relationships of these species: D. erecta is most closely related to D. melanogaster, followed by D. pseudoobscura, D. willistoni and D. littoralis, respectively. Average rates of silent site substitution also do not show a pattern consistent with the accepted phylogeny of these species (Table 2, Figure 1). This is a consequence of the fact that, for comparisons between D. melanogaster and either D. pseudoobscura, D. willistoni or D. littoralis, average rates of silent site substitution exceed an expectation of one substitution per site, indicating that silent sites are 'saturated' in these comparisons. Even so, it is apparent that there may be an increased rate of silent site substitution as well in the D. willistoni lineage. It is unlikely that these results are simply a consequence of an incorrect phylogeny, since the phylogenetic relationships of these species are well established [7].

Our estimate of the average rate of amino-acid substitution per site in known genes between D. melanogaster and D. pseudoobscura (0.071) is nearly the same as previous estimates (0.076) using a different sample of known genes and estimation procedure [33]. In addition, our estimate of the average rate of amino-acid substitution for predicted genes between D. melanogaster and D. erecta (0.071) is similar to that estimated using different methods for a sample of rapidly evolving genes between D. melanogaster and D. yakuba (0.067) [34], a species approximately as divergent from D. melanogaster as D. erecta [35]. Thus the categorical and lineage effects we detect are unlikely to be artifacts of our data or methods. The cause(s) of the increased rate of amino-acid substitution in predicted genes in the D. willistoni lineage remain to be clarified, but are most probably related to increased rates of protein evolution detected previously in the D. saltans lineage [36], which have been explained by a shift in base composition in the common ancestor of the D. saltans and D. willistoni groups (see below, and [37]).

In D. melanogaster, it is well established that coding sequences have a higher GC content, relative to genomic averages, due to the preferential use of codons ending in C or G [38,39]. This pattern holds in the closely related species D. erecta, as well as in the more distantly related species D. pseudoobscura and D. littoralis (Table 3). In contrast, our data show that D. willistoni coding sequences have a higher frequency of AT (53%) base-pairs than GC (47%) base-pairs. This shift in base usage in D. willistoni coding sequences is apparent at the dinucelotide level as well, predominantly affecting those dinucleotides that exclusively contain AT or GC. Non-coding sequences of all non-melanogaster species are AT-rich, as in D. melanogaster [40]; slight shifts towards higher AT frequency are observed in the non-coding sequences of the D. willistoni lineage (Table 3).

The shift in base usage in the D. willistoni lineage is also detected in the pattern of synonymous codon usage (see Supplementary Table 2 in the additional data files). Previous analyses of a limited number of coding sequences revealed a shift away from preferred C-ending codons used in the D. melanogaster lineage, towards T-ending codons in the D. willistoni lineage [7,41]. Our data indicate that this trend holds for a much larger sample of genes (see Supplementary Table 2 in additional data files). For 10 of the 18 amino acids with more than one codon (Arg, Asn, His, Ile, Leu, Lys, Phe, Pro, Thr, Tyr), the most frequently used codon in D. willistoni differs from that in D. melanogaster. All 10 of these changes in synonymous codon usage involve D. willistoni most frequently using an A- or T-ending (or beginning, for example, Leu) codon with D. melanogaster using a G- or C-ending (or beginning) codon, supporting a trend identified originally using only the Adh coding sequence [41]. The most frequently used codon differs between D. melanogaster and D. erecta, D. pseudoobscura and D. littoralis for only two (Asp, Ser), one (Asn) and four (Asn, Ile, Pro, Thr) amino acids, respectively.

Our data also provide an opportunity to study basic features of non-coding conservation in Drosophila. which remain largely unexplored. As shown in Figures 2 and 4, a substantial proportion of non-coding sequences are conserved in Drosophila, especially in pairwise comparisons between D. melanogaster and D. erecta. Levels of conservation appear to plateau in more divergent comparisons, with a tendency for D. pseudoobscura to show higher levels of non-coding conservation relative to D. willistoni or D. littoralis in pairwise comparisons with D. melanogaster. Few, if any, non-coding sequences are conserved between D. melanogaster and A. gambiae (Figure 4, see also [23]). There is also regional variation in levels of non-coding conservation in the Drosophila genome, as illustrated by contrasting conservation between D. melanogaster and D. erecta, for example, in the eve (Figure 2) and ap (Figure 4) regions.

To estimate levels of sequence conservation in non-coding regions and to contrast patterns of coding with non-coding conservation, we aligned genomic sequences using the AVID alignment tool [42]. AVID is a global alignment tool that works by recursively finding co-linear 'anchors' of maximal sequence identity; therefore, locally inverted or transposed sequences that might be conserved will not be included in our analysis. Conserved non-coding sequences (CNCSs), defined as windows of 10 bp or greater with 90% or greater nucleotide identity, were identified in pairwise alignments using the VISTA program [43]. These parameters were chosen to identify short, highly conserved sequences found in Drosophila non-coding regions [44]. We used D. melanogaster as the reference species in pairwise comparisons with non-melanogaster species, and Release 3 annotations [14] exported from Gadfly in VISTA format to classify conserved segments as either coding or non-coding. Transcribed and nontranscribed non-coding sequences were analyzed together, since previous results showed similar patterns of conservation for intergenic and intronic sequences in Drosophila [44].

The results of this analysis are shown in Table 4, which contrasts features of conservation in both coding and non-coding sequences by species. For all species analyzed, coding regions have a higher proportion of sequences that meet our definition of conservation relative to non-coding sequences. In addition, the median segment length surpassing our criterion for conservation is longer for coding sequences relative to non-coding sequences for all species analyzed. These results are expected, as coding sequences are on average thought to experience more intense purifying selection than non-coding sequences [21]. In contrast, the average percent identity of conserved segments is higher for non-coding sequences than coding sequences. This is probably a result of silent site substitution in otherwise functionally constrained coding sequences.

Analysis of levels of conservation by species shows that the increased rate of amino-acid sequence evolution in the D. willistoni lineage detected above may reflect a more widespread phenomenon in the genome of this species. As shown in Table 4, D. willistoni shows unexpectedly low levels of both non-coding and coding conservation, given the accepted phylogeny of the species. These data show that the increased rate of evolution in the D. willistoni lineage is not restricted to coding sequences, rendering coding-sequence-based interpretations of the unusual patterns of molecular evolution in this lineage less tenable (see, for example [7,41]). Together with the changes in base composition in both coding and non-coding sequences noted above, the increased rate of evolution in both coding and non-coding sequences detected in the D. willistoni suggests a genome-wide effect, possibly resulting from a change in mutation pressure or a change in population size at some time during the history of this lineage (see also [37]).

Despite the lineage effect in levels of conservation in the D. willistoni genome, the median length of conserved coding or non-coding segments generally decreases with increasing divergence time as expected (Table 4). However, the average percent identity of conserved coding or non-coding segments identified does not decrease with increasing divergence time. Finally, the ratio of conserved sequences that are coding relative to non-coding increases with increasing divergence time. The ratio of conserved sequences that are coding relative to non-coding is 1.36 for comparisons with D. erecta, but increases to 2.21 for comparisons involving D. pseudoobscura and approximately 3.5 for comparisons involving D. willistoni or D. littoralis.

Changes in the median CNCS length reflect changes in the overall distribution of CNCS lengths in pairwise comparisons between D. melanogaster and either D. erecta, D. pseudoobscura, D. willistoni, or D. littoralis (Figure 5). These data quantitatively describe the pattern of non-coding conservation shown in Figures 2 and 4: CNCS lengths become shorter with increasing divergence but plateau to approximately the same length in the most distant comparisons. The stability of this distribution at more extreme evolutionary distances is apparently insensitive to changes in the proportion of non-coding DNA that is conserved (compare D. willistoni and D. littoralis). Shown for comparison is the distribution of CNCS lengths between D. melanogaster and D. virilis from [44], as well as a reanalysis of this data using the current methods. Differences between the present and previous results for the D. virilis data show the effect of different methods for detecting CNCSs. The differences observed in the distribution of CNCS lengths between the closely related species D. virilis and D. littoralis using the AVID-VISTA method reflect the fact that the D. virilis data were obtained from non-coding regions with known or suspected cis-regulatory function, whereas the data here represent a more random sampling of non-coding regions in the Drosophila genome.

Conservation of non-coding sequences is typically interpreted as evidence of functional constraint and this assumption underlies most phylogenetic footprinting methods. This assumption was questioned by Clark [45], who proposed an alternative hypothesis for non-coding conservation based on heterogeneity in mutation rates (that is, mutational cold spots). To resolve these alternatives we studied the spatial distribution and conservation of spacing between CNCSs in the Drosophila genome. Under a simple mutational cold-spot hypothesis, CNCSs should occur randomly in non-coding DNA and the lengths of 'spacer intervals' between CNCSs should be exponentially distributed [46,47]. In addition, there should be no tendency for the spacing between mutational cold spots to remain conserved between divergent Drosophila species, given the rapid rate of DNA loss in unconstrained sequences in the Drosophila genome [48,49].

As shown in Figure 6 for non-coding comparisons between D. melanogaster and D. pseudoobscura, the frequency spacer interval lengths between CNCSs in the D. melanogaster genome differ significantly from the exponential distribution. The deviation from expected results from an excess of short and long spacer intervals, indicating that CNCSs are clustered in the Drosophila genome. Non-random spacing of CNCSs is also observed in other pairwise species comparisons in the genus Drosophila ([46] and data not shown). In addition, the lengths of homologous spacer intervals are highly correlated across species (Figure 7). This correlation is unlikely to be an artifact of alignment, as the AVID method first aligns regions of local similarity before generating a global alignment. Moreover, similar results have been obtained using non-global alignment methods [46]. These results suggest that spacer interval sequences between CNCSs (and therefore CNCSs themselves) are functionally constrained, and provide evidence against the hypothesis that CNCSs are simply mutational cold spots.

Clusters of CNCSs are readily apparent in VISTA plots of complex gene regions with known cis-regulatory function (Figures 2 and 4). In addition, there is a strong tendency for known cis-regulatory elements to overlap clusters of CNCSs. For example, discrete enhancers that control embryonic expression of eve are contained within discrete CNCS clusters in the region 5' to eve (Figure 2). In contrast, discrete CNCSs clusters are not observed in the region 3' to eve where enhancers overlap one another [50,51]. The correspondence of CNCS clusters and functional enhancers is observed in other regions of the Drosophila genome, such as the discrete muscle-specific enhancer in the fourth intron of ap (Figure 4). The inexact correspondence between enhancer sequences and CNCS clusters is perhaps not unexpected as enhancers are typically defined as the minimal sequence sufficient to recapitulate native expression in a reporter gene assay. Nevertheless, this pattern suggests a functional relationship between cis-regulatory elements and discrete CNCS clusters.

To test the hypothesis that CNCS clusters can predict the location of cis-regulatory elements in the Drosophila genome, we carried out P-element-mediated reporter gene analysis of genomic sequences corresponding to a CNCS cluster in the fourth intron of ap. This CNCS cluster is apparent in pairwise comparisons between D. melanogaster and D. pseudoobscura as well as between D. melanogaster and D. virilis (Figure 4). ap is a LIM-homeobox transcription factor expressed in many tissues in Drosophila, including embryonic expression in the developing brain [52,53]. As shown in Figure 8, the D. melanogaster genomic sequences corresponding to the CNCS cluster in the ap intron 4 drives reporter gene expression in the Drosophila embryo in a specific pattern that recapitulates native ap expression in the developing brain. In addition, when introduced into the genome of D. melanogaster, the homologous fragment from the D. virilis genome also drives reporter gene expression in the same pattern, indicating that the expression pattern resulting from this enhancer has been conserved since the divergence of these two species. Experiments to test the function of CNCS clusters 1, 2, and 3 in the ap region are currently underway.

Although great progress has been made towards understanding the protein-coding component of eukaryotic genome sequences [54,55,56,57], comprehensive genome annotation is far from complete in any metazoan. State-of-the-art statistical and remote-homology gene-prediction methods are successful at identifying the location of exons in unannotated genomic DNA, but are often quite poor at predicting the details of gene structure, necessitating human curation [14]. One of the most useful sources of information for accurately predicting complex gene structures is EST/cDNA data [58]. Predicting the structure of genes for which no EST/cDNA data exists will require alternative approaches, such as comparative gene modeling among divergent species with conserved proteomes in the same group of organisms.

The results of our K_a/K_s analyses presented here give preliminary insight into the prospects of comparative gene modeling using large-scale sequence data in the genus Drosophila. From our findings, we expect that structural details of most Release 3 coding sequences can be verified and improved using pairwise sequence data between divergent species like D. melanogaster and D. pseudoobscura. Our results also indicate that, although it may not be necessary for many genes in the D. melanogaster genome, de novo comparative gene prediction between these species will find the vast majority of as-yet unidentified genes lacking EST/cDNA data. It is important to note that we do not expect to detect all coding exons (especially short exons [20]) in pairwise comparisons, highlighting the added value of multiple species data for comparative exon prediction. In addition, important details of gene models will prove difficult to predict using only comparative data, as amino-acid divergence (especially insertions or deletions) can obscure intron-exon boundaries and other details of gene structure. Moreover, there is inherent uncertainty in the 'correct' gene structure developed from comparative data alone, since two divergent sequences are simultaneously being modeled. Finally, the comparative annotation of UTR sequences awaits the development of methods that accurately predict the non-coding components of gene models.

The patterns of protein-coding sequence evolution detected in our data have important implications for comparative gene prediction. Most notably, the trend we detect for predicted genes to show an increased rate of amino-acid substitution relative to known genes is important, as it may reflect differences in functional constraint or quality of gene models between the two classes of genes. For at least three reasons, we believe that the elevated rate of amino-acid substitution in predicted genes is not a result of poor-quality gene models in this class of genes. First, many of the genes in the predicted class have EST/cDNA data (see Supplementary Table 1), so the details of these gene models are likely to be correct. Second, estimates of K_a (and K_s) are based on aligned sequences; thus gross inaccuracies in gene models that would create gaps in the alignment are excluded from estimates of evolutionary rates. Third, differential rates in these two classes of genes maybe expected, as a high proportion of known genes were selected for study because their mutational inactivation resulted in an obvious phenotype. Thus we favor the interpretation that increased rates of amino-acid substitution in predicted genes results from lower levels of functional constraint.

If this interpretation is correct, our results confirm those of Schmid and co-workers [34,59] who have shown that a large fraction of randomly sampled coding sequences and orphan genes are rapidly evolving in the genus Drosophila. Our results are also consistent with those of Ashburner et al. [60] who show that genes with known mutant phenotypes in D. melanogaster are more likely to have a conserved homolog in GenBank relative to predicted genes with no known phenotype. Similarly, Zdobnov et al. [23] show that D. melanogaster orphan genes tend to exhibit lower levels of conservation in pairwise comparison with A. gambiae [23]. Finally, the interpretation that known and predicted genes differ in their levels of functional constraint is supported by the fact that increased rates of protein evolution in D. willistoni affect predicted genes more strongly than known genes (Table 2). Together these results suggest that there is a large class of functional protein-coding sequences evolving under weak selective constraint in the Drosophila genome [34]. Rates of evolution for this class of genes may be too fast to allow the identification of homologs from extremely divergent species (such as Anopheles) for comparative gene prediction, but slow enough to use comparative data within the genus Drosophila.

Genome rearrangement in Drosophila typically occurs through paracentric inversion, allowing the homolo