Article

Genome sequencing and assembly

We used the previously published sequence and updated assemblies for two Drosophila species, D. melanogaster^{3, 4} (release 4) and D. pseudoobscura⁵ (release 2), and generated DNA sequence data for 10 additional Drosophila genomes by whole-genome shotgun sequencing^{6, 7}. These species were chosen to span a wide variety of evolutionary distances, from closely related pairs such as D. sechellia/D. simulans and D. persimilis/D. pseudoobscura to the distantly related species of the Drosophila and Sophophora subgenera. Whereas the time to the most recent common ancestor of the sequenced species may seem small on an evolutionary timescale, the evolutionary divergence spanned by the genus Drosophila exceeds that of the entire mammalian radiation when generation time is taken into account, as discussed further in ref. 8. We sequenced seven of the new species (D. yakuba, D. erecta, D. ananassae, D. willistoni, D. virilis, D. mojavensis and D. grimshawi) to deep coverage (8.4 to 11.0) to produce high quality draft sequences. We sequenced two species, D. sechellia and D. persimilis, to intermediate coverage (4.9 and 4.1, respectively) under the assumption that the availability of a sister species sequenced to high coverage would obviate the need for deep sequencing without sacrificing draft genome quality. Finally, seven inbred strains of D. simulans were sequenced to low coverage (2.9 coverage from w⁵⁰¹ and 1 coverage of six other strains) to provide population variation data⁹. Further details of the sequencing strategy can be found in Table 1, Supplementary Table 1 and section 1 in Supplementary Information.

Table 1: A summary of sequencing and assembly properties of each new genome

Full table

We generated an initial draft assembly for each species using one of three different whole-genome shotgun assembly programs (Table 1). For D. ananassae, D. erecta, D. grimshawi, D. mojavensis, D. virilis and D. willistoni, we also generated secondary assemblies; reconciliation of these with the primary assemblies resulted in a 7–30% decrease in the estimated number of misassembled regions and a 12–23% increase in the N50 contig size¹⁰ (Supplementary Table 2). For D. yakuba, we generated 52,000 targeted reads across low-quality regions and gaps to improve the assembly. This doubled the mean contig and scaffold sizes and increased the total fraction of high quality bases (quality score (Q) > 40) from 96.5% to 98.5%. We improved the initial 2.9 D. simulans w⁵⁰¹ whole-genome shotgun assembly by filling assembly gaps with contigs and unplaced reads from the 1 assemblies of the six other D. simulans strains, generating a 'mosaic' assembly (Supplementary Table 3). This integration markedly improved the D. simulans assembly: the N50 contig size of the mosaic assembly, for instance, is more than twice that of the initial w⁵⁰¹ assembly (17 kb versus 7 kb).

Finally, one advantage of sequencing genomes of multiple closely related species is that these evolutionary relationships can be exploited to dramatically improve assemblies. D. yakuba and D. simulans contigs and scaffolds were ordered and oriented using pairwise alignment to the well-validated D. melanogaster genome sequence (Supplementary Information section 2). Likewise, the 4–5 D. persimilis and D. sechellia assemblies were improved by assisted assembly using the sister species (D. pseudoobscura and D. simulans, respectively) to validate both alignments between reads and linkage information. For the remaining species, comparative syntenic information, and in some cases linkage information, were also used to pinpoint locations of probable genome mis-assembly, to assign assembly scaffolds to chromosome arms and to infer their order and orientation along euchromatic chromosome arms, supplementing experimental analysis based on known markers (A. Bhutkar, S. Russo, S. Schaeffer, T. F. Smith and W. M. Gelbart, personal communication) (Supplementary Information section 2).

The mitochondrial (mt)DNA of D. melanogaster, D. sechellia, D. simulans (siII), D. mauritiana (maII) and D. yakuba have been previously sequenced^{11, 12}. For the remaining species (except D. pseudoobscura, the DNA from which was prepared from embryonic nuclei), we were able to assemble full mitochondrial genomes, excluding the A+T-rich control region (Supplementary Information section 2)¹³. In addition, the genome sequences of three Wolbachia endosymbionts (Wolbachia wSim, Wolbachia wAna and Wolbachia wWil) were assembled from trace archives, in D. simulans, D. ananassae and D. willistoni, respectively¹⁴. All of the genome sequences described here are available in FlyBase (www.flybase.org) and GenBank (www.ncbi.nlm.nih.gov) (Supplementary Tables 4 and 5).

Repeat and transposable element annotation

Repetitive DNA sequences such as transposable elements pose challenges for whole-genome shotgun assembly and annotation. Because the best approach to transposable element discovery and identification is still an active and unresolved research question, we used several repeat libraries and computational strategies to estimate the transposable element/repeat content of the 12 Drosophila genome assemblies (Supplementary Information section 3). Previously curated transposable element libraries in D. melanogaster provided the starting point for our analysis; to limit the effects of ascertainment bias, we also developed de novo repeat libraries using PILER-DF^{15, 16} and ReAS¹⁷. We used four transposable element/repeat detection methods (RepeatMasker, BLASTER-TX, RepeatRunner and CompTE) in conjunction with these transposable element libraries to identify repetitive elements in non-melanogaster species. We assessed the accuracy of each method by calibration with the estimated 5.5% transposable element content in the D. melanogaster genome, which is based on a high-resolution transposable element annotation¹⁸ (Supplementary Fig. 1). On the basis of our results, we suggest a hybrid strategy for new genome sequences, employing translated BLAST with general transposable element libraries and RepeatMasker with species-specific ReAS libraries to estimate the upper and lower bound on transposable element content.

Protein-coding gene annotation

We annotated protein-coding sequences in the 11 non-melanogaster genomes, using four different de novo gene predictors (GeneID¹⁹, SNAP²⁰, N-SCAN²¹ and CONTRAST²²); three homology-based predictors that transfer annotations from D. melanogaster (GeneWise²³, Exonerate²⁴, GeneMapper²⁵); and one predictor that combined de novo and homology-based evidence (Gnomon²⁶). These gene prediction sets were combined using GLEAN, a gene model combiner that chooses the most probable combination of start, stop, donor and acceptor sites from the input predictions^{27, 28}. All analyses reported here, unless otherwise noted, relied on a reconciled consensus set of predicted gene models—the GLEAN-R set (Table 2, and Supplementary Information section 4.1).

Table 2: A summary of annotated features across all 12 genomes

Full table

Quality of gene models

As the first step in assessing the quality of the GLEAN-R gene models, we used expression data from microarray experiments on adult flies, with arrays custom-designed for D. simulans, D. yakuba, D. ananassae, D. pseudoobscura, D. virilis and D. mojavensis²⁹ (GEO series GSE6640; Supplementary Information section 4.2). We detected expression significantly above negative controls (false-discovery-rate-corrected Mann–Whitney U (MWU) P < 0.001) for 77–93% of assayed GLEAN-R models, representing 50–68% of the total GLEAN-R predictions in each species (Supplementary Table 6). Evolutionarily conserved gene models are much more likely to be expressed than lineage-specific ones (Fig. 2). Although these data cannot confirm the detailed structure of gene models, they do suggest that the majority of GLEAN-R models contain sequence that is part of a poly-adenylated transcript. Approximately 20% of transcription in D. melanogaster seems to be unassociated with protein-coding genes³⁰, and our microarray experiments fail to detect conditionally expressed genes. Thus, transcript abundance cannot conclusively establish the presence or absence of a protein-coding gene. Nonetheless, we believe these expression data increase our confidence in the reliability of the GLEAN-R models, particularly those supported by homology evidence (Fig. 2).

Figure 2: Gene models in 12 Drosophila genomes.

Number of gene models that fall into one of five homology classes: single-copy orthologues in all species (single-copy orthologues), conserved in all species as orthologues or paralogues (conserved homologues), a D. melanogaster homologue, but not found in all species (patchy homologues with mel.), conserved in at least two species but without a D. melanogaster homologue (patchy homologues, no mel.), and found only in a single lineage (lineage specific). For those species with expression data²⁹, pie charts indicate the fraction of genes in each homology class that fall into one of four evidence classes (see text for details).

High resolution image and legend (142K)

Because the GLEAN-R gene models were built using assemblies that were not repeat masked, it is likely that some proportion of gene models are false positives corresponding to coding sequences of transposable elements. We used RepeatMasker with de novo ReAS libraries and PFAM structural annotations of the GLEAN-R gene set to flag potentially transposable element-contaminated gene models (Supplementary Information section 4.2). These procedures suggest that 5.6–32.3% of gene models in non-melanogaster species correspond to protein-coding content derived from transposable elements (Supplementary Table 7); these transposable element-contaminated gene models are almost exclusively confined to gene predictions without strong homology support (Fig. 2). Transposable element-contaminated gene models are excluded from the final gene prediction set used for subsequent analysis, unless otherwise noted.

Homology assignment

Two independent approaches were used to assign orthology and paralogy relationships among euchromatic D. melanogaster gene models and GLEAN-R predictions. The first approach was a fuzzy reciprocal BLAST (FRB) algorithm, which is an extension of the reciprocal BLAST method³¹ applicable to multiple species simultaneously (Supplementary Information section 5.1). Because the FRB algorithm does not integrate syntenic information, we also used a second approach based on Synpipe (Supplementary Information section 5.2), a tool for synteny-aided orthology assignment³². To generate a reconciled set of homology calls, pairwise Synpipe calls (between each species and D. melanogaster) were mapped to GLEAN-R models, filtered to retain only 1:1 relationships, and added to the FRB calls when they did not conflict and were non-redundant. This reconciled FRB + Synpipe set of homology calls forms the basis of our subsequent analyses. There were 8,563 genes with single-copy orthologues in the melanogaster group and 6,698 genes with single-copy orthologues in all 12 species; similar numbers of genes were also obtained with an independent approach³³. Most single-copy orthologues are expressed and are free from potential transposable element contamination, suggesting that the reconciled orthologue set contains robust and high-quality gene models (Fig. 2).

Validation of homology calls

Because both the FRB algorithm and Synpipe rely on BLAST-based methods to infer similarities, rapidly evolving genes may be overlooked. Moreover, assembly gaps and poor-quality sequence may lead to erroneous inferences of gene loss. To validate putative gene absences, we used a synteny-based GeneWise pipeline to find potentially missed homologues of D. melanogaster proteins (Supplementary Information section 5.4). Of the 21,928 cases in which a D. melanogaster gene was absent from another species in the initial homology call set, we identified plausible homologues for 13,265 (60.5%), confirmed 4,546 (20.7%) as genuine absences, and were unable to resolve 4,117 (18.8%). Because this approach is conservative and only confirms strongly supported absences, we are probably underestimating the number of genuine absences.

Coding gene alignment and filtering

Investigating the molecular evolution of orthologous and paralogous genes requires accurate multi-species alignments. Initial amino acid alignments were generated using TCOFFEE³⁴ and converted to nucleotide alignments (Supplementary Table 8). To reduce biases in downstream analyses, a simple computational screen was developed to identify and mask problematic regions of each alignment (Supplementary Information section 6). Overall, 2.8% of bases were masked in the melanogaster group alignments, and 3.0% of bases were masked in the full 12 species alignments, representing 8.5% and 13.8% of alignment columns, respectively. The vast majority of masked bases are masked in no more than one species (Supplementary Fig. 3), suggesting that the masking procedure is not simply eliminating rapidly evolving regions of the genome. We find an appreciably higher frequency of masked bases in lower-quality D. simulans and D. sechellia assemblies, compared to the more divergent (from D. melanogaster) but higher-quality D. erecta and D. yakuba assemblies, suggesting a higher error rate in accurately predicting and aligning gene models in lower-quality assemblies (Supplementary Information section 6 and Supplementary Fig. 3). We used masked versions of the alignments, including only the longest D. melanogaster transcripts for all subsequent analysis unless otherwise noted.

Annotation of non-coding (nc)RNA genes

Using de novo and homology-based approaches we annotated over 9,000 ncRNA genes from recognized ncRNA classes (Table 2, and Supplementary Information section 7). In contrast to the large number of predictions observed for many ncRNA families in vertebrates (due in part to large numbers of ncRNA pseudogenes^{35, 36}), the number of ncRNA genes per family predicted by RFAM and tRNAscan in Drosophila is relatively low (Table 2). This suggests that ncRNA pseudogenes are largely absent from Drosophila genomes, which is consistent with the low number of protein-coding pseudogenes in Drosophila³⁷. The relatively low numbers of some classes of ncRNA genes (for example, small nucleolar (sno)RNAs) in the Drosophila subgenus are likely to be an artefact of rapid rates of evolution in these types of genes and the limitation of the homology-based methods used to annotate distantly related species.

Coarse-level similarities among Drosophilids

At a coarse level, genome structure is well conserved across the 12 sequenced species. Total genome size estimated by flow cytometry varies less than threefold across the phylogeny, ranging from 130 Mb (D. mojavensis) to 364 Mb (D. virilis)³⁸ (Table 2), in contrast to the order of magnitude difference between Drosophila and mammals. Total protein-coding sequence ranges from 38.9 Mb in D. melanogaster to 65.4 Mb in D. willistoni. Intronic DNA content is also largely conserved, ranging from 19.6 Mb in D. simulans to 24.0 Mb in D. pseudoobscura (Table 2). This contrasts dramatically with transposable element-derived genomic DNA content, which varies considerably across genomes (Table 2) and correlates significantly with euchromatic genome size (estimated as the summed length of contigs > 200 kb) (Kendall's = 0.70, P = 0.0016).

To investigate overall conservation of genome architecture at an intermediate scale, we analysed synteny relationships across species using Synpipe³² (Supplementary Information section 9.1). Synteny block size and average number of genes per block varies across the phylogeny as expected, with the number of blocks increasing and the average size of blocks decreasing with increasing evolutionary distance from D. melanogaster (A. Bhutkar, S. Russo, T. F. Smith and W. M. Gelbart, personal communication) (Supplementary Fig. 4). We inferred 112 syntenic blocks between D. melanogaster and D. sechellia (with an average of 122 genes per block), compared to 1,406 syntenic blocks between D. melanogaster and D. grimshawi (with an average of 8 genes per block). On average, 66% of each genome assembly was covered by syntenic blocks, ranging from 68% in D. sechellia to 58% in D. grimshawi.

Similarity across genomes is largely recapitulated at the level of individual genes, with roughly comparable numbers of predicted protein-coding genes across the 12 species (Table 2). The majority of predicted genes in each species have homologues in D. melanogaster (Table 2, Supplementary Table 9). Moreover, most of the 13,733 protein-coding genes in D. melanogaster are conserved across the entire phylogeny: 77% have identifiable homologues in all 12 genomes, 62% can be identified as single-copy orthologues in the six genomes of the melanogaster group and 49% can be identified as single-copy orthologues in all 12 genomes. The number of functional non-coding RNA genes predicted in each Drosophila genome is also largely conserved, ranging from 584 in D. mojavensis to 908 in D. ananassae (Table 2).

There are several possible explanations for the observed interspecific variation in gene content. First, approximately 700 D. melanogaster gene models have been newly annotated since the FlyBase Release 4.3 annotations used in the current study, reducing the discrepancy between D. melanogaster and the other sequenced genomes in this study. Second, because low-coverage genomes tend to have more predicted gene models, we suspect that artefactual duplication of genomic segments due to assembly errors inflates the number of predicted genes in some species. Finally, the non-melanogaster species have many more predicted lineage-specific genes than D. melanogaster, and it is possible that some of these are artefactual. In the absence of experimental evidence, it is difficult to distinguish genuine lineage-specific genes from putative artefacts. Future experimental work will be required to fully disentangle the causes of interspecific variation in gene number.

Abundant genome rearrangements during Drosophila evolution

To study the structural relationships among genomes on a finer scale, we analysed gene-level synteny between species pairs. These synteny maps allowed us to infer the history and locations of fixed genomic rearrangements between species. Although Drosophila species vary in their number of chromosomes, there are six fundamental chromosome arms common to all species. For ease of denoting chromosomal homology, these six arms are referred to as 'Muller elements' after Hermann J. Muller, and are denoted A–F. Although most pairs of orthologous genes are found on the same Muller element, there is extensive gene shuffling within Muller elements between even moderately diverged genomes (Fig. 3, and Supplementary Information section 9.1).

Figure 3: Synteny plots for Muller elements B and C with respect to D. melanogaster gene order.

The horizontal axis shows D. melanogaster gene order for Muller elements B and C, and the vertical axis maps homologous locations^{32, 155} in individual species (a–f in increasing evolutionary distance from D. melanogaster). Left to right on the x axis is from telomere to centromere for Muller element B, followed by Muller element C from centromere to telomere. Red and green lines represent syntenic segments in the same or reverse orientation along the chromosome relative to D. melanogaster, respectively. Blue segments show gene transposition of genes from one element to the other.

High resolution image and legend (145K)

Previous analysis has revealed heterogeneity in rearrangement rates among close relatives: careful inspection of 29 inversions that differentiate the chromosomes of D. melanogaster and D. yakuba revealed that 28 were fixed in the lineage leading to D. yakuba, and only one was fixed on the lineage leading to D. melanogaster³⁹. Rearrangement rates are also heterogeneous across the genome among the 12 species: simulations reject a random-breakage model, which assumes that all sites are free to break in inversion events, but fail to reject a model of coldspots and hotspots for breakpoints (S. Schaeffer, personal communication). Furthermore, inversions seem to have played important roles in the process of speciation in at least some of these taxa⁴⁰.

One particularly striking example of the dynamic nature of genome micro-structure in Drosophila is the homeotic homeobox (Hox) gene cluster(s)⁴¹. Hox genes typically occur in genomic clusters, and this clustering is conserved across many vertebrate and invertebrate taxa, suggesting a functional role for the precise and collinear arrangement of these genes. However, several cluster splits have been previously identified in Drosophila^{42, 43}, and the 12 Drosophila genome sequences provide additional evidence against the functional importance of Hox gene clustering in Drosophila. There are seven different gene arrangements found across 13 Drosophila species (the 12 sequenced genomes and D. buzzatii), with no species retaining the inferred ancestral gene order⁴⁴. It thus seems that, in Drosophila, Hox genes do not require clustering to maintain proper function, and are a powerful illustration of the dynamism of genome structure across the sequenced genomes.

Transposable element evolution

Mobile, repetitive transposable element sequences are a particularly dynamic component of eukaryotic genomes. Transposable element/repeat content (in scaffolds >200 kb) varies by over an order of magnitude across the genus, ranging from 2.7% in D. simulans and D. grimshawi to 25% in D. ananassae (Table 2, and Supplementary Fig. 1). These data support the lower euchromatic transposable element content in D. simulans relative to D. melanogaster⁴⁵, and reveal that euchromatic transposable element/repeat content is generally similar within the melanogaster subgroup. Within the Drosophila subgenus, D. grimshawi has the lowest transposable element/repeat content, possibly relating to its ecological status as an island endemic, which may minimize the chance for horizontal transfer of transposable element families. Finally, the highest levels of transposable element/repeat content are found in D. ananassae and D. willistoni. These species also have the highest numbers of pseudo-transfer (t)RNA genes (Table 2), indicating a potential relationship between pseudo-tRNA genesis and repetitive DNA, as has been established in the mouse genome³⁶.

Different classes of transposable elements can vary in abundance owing to a variety of host factors, motivating an analysis of the intragenomic ecology of transposable elements in the 12 genomes. In D. melanogaster, long terminal repeat (LTR) retrotransposons have the highest abundance, followed by LINE (long interspersed nuclear element)-like retrotransposons and terminal inverted repeat (TIR) DNA-based transposons¹⁸. An unbiased, conservative approach (Supplementary Information section 3) for estimating the rank order abundance of major transposable element classes suggests that these abundance trends are conserved across the entire genus (Supplementary Fig. 5). Two exceptions are an increased abundance of TIR elements in D. erecta and a decreased abundance of LTR elements in D. pseudoobscura; the latter observation may represent an assembly artefact because the sister species D. persimilis shows typical LTR abundance. Given that individual instances of transposable element repeats and transposable element families themselves are not conserved across the genus, the stability of abundance trends for different classes of transposable elements is striking and suggests common mechanisms for host–transposable element co-evolution in Drosophila.

Although comprehensive analysis of the structural and evolutionary relationships among families of transposable elements in the 12 genomes remains a major challenge for Drosophila genomics, some initial insights can be gleaned from analysis of particularly well-characterized transposable element families. Previous analysis has shown variable dynamics for the most abundant transposable element family (DINE-1)⁴⁶ in the D. melanogaster genome^{18, 47}: although inactive in D. melanogaster⁴⁸, DINE-1 has experienced a recent transpositional burst in D. yakuba⁴⁹. Our analysis confirms that this element is highly abundant in all of the other sequenced genomes of Drosophila, but is not found outside of Diptera^{50, 51}. Moreover, the inferred phylogenetic relationship of DINE-1 paralogues from several Drosophila species suggests vertical transmission as the major mechanism for DINE-1 propagation. Likewise, analysis of the Galileo and 1360 transposons reveals a widespread but discontinuous phylogenetic distribution for both families, notably with both families absent in the geographically isolated Hawaiian species, D. grimshawi⁵². These results are consistent with an ancient origin of the Galileo and 1360 families in the genus and subsequent horizontal transfer and/or loss in some lineages.

The use of these 12 genomes also facilitated the discovery of transposable element lineages not yet documented in Drosophila, specifically the P instability factor (PIF) superfamily of DNA transposons. Our analysis indicates that there are four distinct lineages of this transposon in Drosophila, and that this element has indeed colonized many of the sequenced genomes⁵³. This superfamily is particularly intriguing given that PIF-transposase-like genes have been implicated in the origin of at least seven different genes during the Drosophila radiation⁵³, suggesting that not only do transposable elements affect the evolution of genome structure, but that their domestication can play a part in the emergence of novel genes.

D. melanogaster maintains its telomeres by occasional targeted transposition of three telomere-specific non-LTR retrotransposons (HeT-A, TART and TAHRE) to chromosome ends^{54, 55} and not by the more common mechanism of telomerase-generated G-rich repeats⁵⁶. Multiple telomeric retrotransposons have originated within the genus, where they now maintain telomeres, and recurrent loss of most of the ORF2 from telomeric retrotransposons (for example, TAHRE) has given rise to half-telomeric-retrotransposons (for example, HeT-A) during Drosophila evolution⁵⁷. The phylogenetic relationship among these telomeric elements is congruent with the species phylogeny, suggesting that they have been vertically transmitted from a common ancestor⁵⁷.

ncRNA gene family evolution

Using ncRNA gene annotations across the 12-species phylogeny, we inferred patterns of gene copy number evolution in several ncRNA families. Transfer RNA genes are the most abundant family of ncRNA genes in all 12 genomes, with 297 tRNAs in D. melanogaster and 261–484 tRNA genes in the other species (Table 2). Each genome encodes a single selenocysteine tRNA, with the exception of D. willistoni, which seems to lack this gene (R. Guigo, personal communication). Elevated tRNA gene counts in D. ananassae and D. willistoni are explained almost entirely by pseudo-tRNA gene predictions. We infer from the lack of pseudo-tRNAs in most Drosophila species, and from similar numbers of tRNAs obtained from an analysis of the chicken genome (n = 280)⁵⁸, that the minimal metazoan tRNA set is encoded by 300 genes, in contrast to previous estimates of 497 in human and 659 in Caenorhabditis elegans^{59, 60}. Similar numbers of snoRNAs are predicted in the D. melanogaster subgroup (n = 242–255), in which sequence similarity is high enough for annotation by homology, with fewer snoRNAs (n = 194–216) annotated in more distant members of the Sophophora subgenus, and even fewer snoRNAs (n = 139–165) predicted in the Drosophila subgenus, in which annotation by homology becomes much more difficult.

Of 78 previously reported micro (mi)RNA genes, 71 (91%) are highly conserved across the entire genus, with the remaining seven genes (mir-2b-1, -289, -303, -310, -311, -312 and -313) restricted to the subgenus Sophophora (Supplementary Information section 7.2). All the species contain similar numbers of spliceosomal snRNA genes (Table 2), including at least one copy each of the four U12-dependent (minor) spliceosomal RNAs, despite evidence for birth and death of these genes and the absence of stable subtypes⁶¹. The unusual, lineage-specific expansion in size of U11 snRNA, previously described in Drosophila^{61, 62}, is even more extreme in D. willistoni. We annotated 99 copies of the 5S ribosomal (r)RNA gene in a cluster in D. melanogaster, and between 13 and 73 partial 5S rRNA genes in clusters in the other genomes. Finally, we identified members of several other classes of ncRNA genes, including the RNA components of the RNase P (1 per genome) and the signal recognition particle (SRP) RNA complexes (1–3 per genome), suggesting that these functional RNAs are involved in similar biological processes throughout the genus. We were only able to locate the roX (RNA on X)^{63, 64} genes involved in dosage compensation using nucleotide homology in the melanogaster subgroup, although analyses incorporating structural information have identified roX genes in other members of the genus⁶⁵.

We investigated the evolution of rRNA genes in the 12 sequenced genomes, using trace archives to locate sequence variants within the transcribed portions of these genes. This analysis revealed moderate levels of variation that are not distributed evenly across the rRNA genes, with fewest variants in conserved core coding regions, more variants in coding expansion regions, and higher still variant abundances in non-coding regions. The level and distribution of sequence variation in rRNA genes are suggestive of concerted evolution, in which recombination events uniformly distribute variants throughout the rDNA loci, and selection dictates the frequency to which variants can expand⁶⁶.

Protein-coding gene family evolution

For a general perspective on how the protein-coding composition of these 12 genomes has changed, we examined gene family expansions and contractions in the 11,434 gene families (including those of size one in each species) predicted to be present in the most recent common ancestor of the two subgenera. We applied a maximum likelihood model of gene gain and loss⁶⁷ to estimate rates of gene turnover. This analysis suggests that gene families expand or contract at a rate of 0.0012 gains and losses per gene per million years, or roughly one fixed gene gain/loss across the genome every 60,000 yr⁶⁸. Many gene families (4,692 or 41.0%) changed in size in at least one species, and 342 families showed significantly elevated (P < 0.0001) rates of gene gain and loss compared to the genomic average, indicating that non-neutral processes may play a part in gene family evolution. Twenty-two families exhibit rapid copy number evolution along the branch leading to D. melanogaster (eighteen contractions and four expansions; Supplementary Table 10). The most common Gene Ontology (GO) terms among families with elevated rates of gain/loss include 'defence response', 'protein binding', 'zinc ion binding', 'proteolysis', and 'trypsin activity'. Interestingly, genes involved in 'defence response' and 'proteolysis' also show high rates of protein evolution (see below). We also found heterogeneity in overall rates of gene gain and loss across lineages, although much of this variation could result from interspecific differences in assembly quality⁶⁸.

Lineage-specific genes

The vast majority of D. melanogaster proteins that can be unambiguously assigned a homology pattern (Supplementary Information section 5) are inferred to be ancestrally present at the genus root (11,348/11,644, or 97.5%). Of the 296 non-ancestrally present genes, 252 are either Sophophora-specific, or have a complicated pattern of homology requiring more than one gain and/or loss on the phylogeny, and are not discussed further. The remaining 44 proteins include 14 present in the melanogaster group, 23 present only in the melanogaster subgroup, 3 unique to the melanogaster species complex, and 4 found in D. melanogaster only. Because we restricted this analysis to unambiguous homologues of high-confidence protein-coding genes in D. melanogaster⁸, we are probably undercounting the number of genes that have arisen de novo in any particular lineage. However, ancestrally heterochromatic genes that are currently euchromatic in D. melanogaster may spuriously seem to be lineage-specific.

The 44 lineage-specific genes (Supplementary Table 11) differ from ancestrally present genes in several ways. They have a shorter median predicted protein length (lineage-specific median 177 amino acids, other median 421 amino acids, MWU, P = 3.6  10^-13), are more likely to be intronless (Fisher's exact test (FET), P = 6.2  10^-6), and are more likely to be located in the intron of another gene on the opposite strand (FET, P = 3.5  10^-4). In addition, 18 of these 44 genes are testis- or accessory-gland-specific in D. melanogaster, a significantly greater fraction than is found in the ancestral set (FET, P = 1.25  10^-4). This is consistent with previous observations that novel genes are often testis-specific in Drosophila^{69, 70, 71, 72, 73} and expression studies on seven of the species show that species-restricted genes are more likely to exhibit male-biased expression²⁹. Further, these genes are significantly more tissue-specific in expression (as measured by ; ref. 74) (MWU, P = 9.6  10^-6), and this pattern is not solely driven by genes with testis-specific expression patterns.

Positive selection and selective constraints in Drosophila genomes

To study the molecular evolution of protein-coding genes, we estimated rates of synonymous and non-synonymous substitution in 8,510 single-copy orthologues within the six melanogaster group species using PAML⁷⁵ (Supplementary Information section 11.1); synonymous site saturation prevents analysis of more divergent comparisons. We investigate only single-copy orthologues because when paralogues are included, alignments become increasingly problematic. Rates of amino acid divergence for single-copy orthologues in all 12 species were also calculated; these results are largely consistent with the analysis of non-synonymous divergence in the melanogaster group, and are not discussed further.

To understand global patterns of divergence and constraint across functional classes of genes, we examined the distributions of ( = d_N/d_S, the ratio of non-synonymous to synonymous divergence) across Gene Ontology categories (GO)⁷⁶, excluding GO annotations based solely on electronic support (Supplementary Information section 11.2). Most functional categories of genes are strongly constrained, with median estimates of much less than one. In general, functionally similar genes are similarly constrained: 31.8% of GO categories have significantly lower variance in than expected (q-value true-positive test⁷⁷). Only 11% of GO categories had statistically significantly elevated (relative to the median of all genes with GO annotations) at a 5% false-discovery rate (FDR), suggesting either positive selection or a reduction in selective constraint. The GO categories with elevated include the biological process terms 'defence response', 'proteolysis', 'DNA metabolic process' and 'response to biotic stimulus'; the molecular function terms 'transcription factor activity', 'peptidase activity', 'receptor binding', 'odorant binding', 'DNA binding', 'receptor activity' and 'G-protein-coupled receptor activity'; and the cellular location term 'extracellular' (Fig. 4, and Supplementary Table 12). Similar results are obtained when d_N is compared across GO categories, suggesting that in most cases differences in among GO categories is driven by amino acid rather than synonymous site substitutions. The two exceptions are the molecular function terms 'transcription factor activity' and 'DNA binding activity', for which we observe significantly decelerated d_S (FDR = 7.2  10^-4 for both; Supplementary Information section 11.2) and no significant differences in d_N.

Figure 4: Patterns of constraint and positive selection among GO terms.

Distribution of average per gene and the negative log₁₀ of the probability of positive selection (Supplementary Information section 11.2) for genes annotated with: a, biological process GO terms; b, cellular component GO terms; and c, molecular function GO terms. Only GO terms with 200 or more genes annotated are plotted. See Supplementary Table 12 for median values and significance. Note that most genes evolve under evolutionary constraint at most of their sites, leading to low values of ; even genes that experience positive selection do not typically have an average across all codons that exceeds one.

High resolution image and legend (170K)

To distinguish possible positive selection from relaxed constraint, we tested explicitly for genes that have a subset of codons with signatures of positive selection, using codon-based likelihood models of molecular evolution, implemented in PAML^{78, 79} (Supplementary Information section 11.1). Although this test is typically regarded as a conservative test for positive selection, it may be confounded by selection at synonymous sites. However, selection at synonymous sites (that is, codon bias, see below) is quite weak. Moreover, variability in presented here tends to reflect variability in d_N. We therefore believe that it is appropriate to treat synonymous sites as nearly neutral and sites with  > 1 as consistent with positive selection. Despite a number of functional categories with evidence for elevated , 'helicase activity' is the only functional category significantly more likely to be positively selected (permutation test, P = 2  10^-4, FDR = 0.007; Supplementary Table 12); the biological significance of this finding merits further investigation. Furthermore, within each GO class, there is greater dispersion among genes in their probability of positive selection than in their estimate of (MWU one-tailed, P = 0.011; Supplementary Information section 11.1), suggesting that although functionally similar genes share patterns of constraint, they do not necessarily show similar patterns of positive selection (Fig. 4).

Interestingly, protein-coding genes with no annotated ('unknown') function in the GO database seem to be less constrained (permutation test, P < 1  10^-4, FDR = 0.006)⁸⁰ and to have on average lower P-values for the test of positive selection than genes with annotated functions (permutation test, P = 0.001, FDR = 0.058). It is unlikely that this observation results entirely from an over-representation of mis-annotated or non-protein-coding genes in the 'unknown' functional class, because this finding is robust to the removal of all D. melanogaster genes predicted to be non-protein-coding in ref. 8. The bias in the way biological function is ascribed to genes (to laboratory-induced, easily scorable functions) leaves open the possibility that unannotated biological functions may have an important role in evolution. Indeed, genes with characterized mutant alleles in FlyBase evolve significantly more slowly than other genes (median _{with alleles} = 0.0525 and _{without alleles} = 0.0701; MWU, P < 1  10^-16).

Previous work has suggested that a substantial fraction of non-synonymous substitutions in Drosophila were fixed through positive selection^{81, 82, 83, 84, 85}. We estimate that 33.1% of single-copy orthologues in the melanogaster group have experienced positive selection on at least a subset of codons (q-value true-positive tests⁷⁷) (Supplementary Information section 11.1). This may be an underestimate, because we have only examined single-copy orthologues, owing to difficulties in producing accurate alignments of paralogues by automated methods. On the basis of the 878 genes inferred to have experienced positive selection with high confidence (FDR < 10%), we estimated that an average of 2% of codons in positively selected genes have  > 1. Thus, several lines of evidence, based on different methodologies, suggest that patterns of amino acid fixation in Drosophila genomes have been shaped extensively by positive selection.

The presence of functional domains within a protein may lead to heterogeneity in patterns of constraint and adaptation along its length. Among genes inferred to be evolving by positive selection at a 10% FDR, 63.7% (q-value true-positive tests⁷⁷) show evidence for spatial clustering of positively selected codons (Supplementary Information section 11.2). Spatial heterogeneity in constraint is further supported by contrasting for codons inside versus outside defined InterPro domains (genes lacking InterPro domains are treated as 'outside' a defined InterPro domain). Codons within InterPro domains were significantly more conserved than codons outside InterPro domains (median : 0.062 InterPro domains, 0.084 outside InterPro domains; MWU, P < 2.2  10^-16; Supplementary Information section 11.2). Similarly, there were significantly more positively selected codons outside of InterPro domains than inside domains (FET P < 2.2 10^-16), suggesting that in addition to being more constrained, codons in protein domains are less likely to be targets of positive selection (Supplementary Fig. 6).

Factors affecting the rate of protein evolution in Drosophila

The sequenced genomes of the melanogaster group provide unprecedented statistical power to identify factors affecting rates of protein evolution. Previous analyses have suggested that although the level of gene expression consistently seems to be a major determinant of variation in rates of evolution among proteins^{86, 87}, other factors probably play a significant, if perhaps minor, part^{88, 89, 90, 91}. In Drosophila, although highly expressed genes do evolve more slowly, breadth of expression across tissues, gene essentiality and intron number all also independently correlate with rates of protein evolution, suggesting that the additional complexities of multicellular organisms are important factors in modulating rates of protein evolution⁷⁸. The presence of repetitive amino acid sequences has a role as well: non-repeat regions in proteins containing repeats evolve faster and show more evidence for positive selection than genes lacking repeats⁹².

These data also provide a unique opportunity to examine the impact of chromosomal location on evolutionary rates. Population genetic theory predicts that for new recessive mutations, both purifying and positive selection will be more efficient on the X chromosome given its hemizygosity in males⁹³. In contrast, the lack of recombination on the small, mainly heterochromatic dot chromosome^{94, 95} is expected to reduce the efficacy of selection⁹⁶. Because codon bias, or the unequal usage of synonymous codons in protein-coding sequences, reflects weak but pervasive selection, it is a sensitive metric for evaluating the efficacy of purifying selection. Consistent with expectation, in all 12 species, we find significantly elevated levels of codon bias on the X chromosome and significantly reduced levels of codon bias on the dot chromosome⁹⁷. Furthermore, X-chromosome-linked genes are marginally over-represented within the set of positively selected genes in the melanogaster group (FET, P = 0.055), which is consistent with increased rates of adaptive substitution on this chromosome. This analysis suggests that chromosomal context also serves to modulate rates of molecular evolution in protein-coding genes.

To examine further the impact of genomic location on protein evolution, we examined the subset of genes that have moved within or between chromosome arms^{32, 98}. Genes inferred to have moved between Muller elements have a significantly higher rate of protein evolution than genes inferred to have moved within a Muller element (MWU, P = 1.32 10^-14) and genes that have maintained their genomic position (MWU, P = 0.008) (Supplementary Fig. 7). Interestingly, genes that move within Muller elements have a significantly lower rate of protein evolution than those for which genomic locations have been maintained (MWU, P = 3.85 10^-14). It remains unclear whether these differences reflect underlying biases in the types of genes that move inter- versus intra-chromosomally, or whether they are due to in situ patterns of evolution in novel genomic contexts.

Codon bias

Codon bias is thought to enhance the efficiency and/or accuracy of translation^{99, 100, 101} and seems to be maintained by mutation–selection–drift balance^{101, 102, 103, 104}. Across the 12 Drosophila genomes, there is more codon bias in the Sophophora subgenus than in the Drosophila subgenus, and a previously noted^{105, 106, 107, 108, 109} striking reduction in codon bias in D. willistoni^{110, 111} (Fig. 5). However, with only minor exceptions, codon preferences for each amino acid seem to be conserved across 11 of the 12 species. The striking exception is D. willistoni, in which codon usage for 6 of 18 redundant amino acids has diverged (Fig. 5). Mutation alone is not sufficient to explain codon-usage bias in D. willistoni, which is suggestive of a lineage-specific shift in codon preferences^{111, 112}. We found evidence for a lineage-specific genomic reduction in codon bias in D. melanogaster (Fig. 5), as has been suggested previously^{113, 114, 115, 116, 117, 118, 119}. In addition, maximum-likelihood estimation of the strength of selection on synonymous sites in 8,510 melanogaster group single-copy orthologues revealed a marked reduction in the number of genes under selection for increased codon bias in D. melanogaster relative to its sister species D. sechellia¹²⁰.

Figure 5: Deviations in codon bias from D. melanogaster in 11 Drosophila species.

The upper panel depicts differences in ENC (effective number of codons) between D. melanogaster and the 11 non-melanogaster species, calculated on a gene-by-gene basis. Note that increasing levels of ENC indicates a decrease in codon bias. The Sophophora subgenus in general has higher levels of codon bias than the Drosophila subgenus with the exception of D. willistoni, which shows a dramatic reduction in codon bias. The lower panel shows the 7 codons for which preference changes across the 12 Drosophila species. A dot indicates identical codon preference to D. melanogaster; otherwise the preferred codon is indicated.

High resolution image and legend (96K)