A total of 2,425 protein-coding genes were retained in the ancestor, out of a total of 4,351 genes in E. coli. Under the procedures for reconstructing the ancestral genome (A in Figure 1), a requirement for inclusion in the ancestor was presence in E. coli MG1655. As a result, the reconstructed genome is expected to be smaller than that of the real ancestor, because it does not include genes that were present in the ancestor but lost independently from both E. coli and Buchnera lineages. Despite the removal of 44% of the E. coli genes, over 99% of genes present in Buchnera were also retained in the reconstructed ancestor. This outcome strongly supports this parsimony approach to reconstructing the gene inventory of the shared ancestor of E. coli and Buchnera, and also of the hypothesis that Buchnera evolved from such an ancestor through gene loss. Only two of the 560 protein-coding genes that show clear orthology between E. coli and Buchnera would have been removed (due to absence from both Vibrio cholerae and Yersinia pestis). The two genes (dnaC and dnaT) are linked in both Buchnera and E. coli and were retained in the ancestor.

Fragments for which the order and orientation of genes are the same in Buchnera and E. coli were assumed to be present in the common ancestor. A syntenic fragment was recognized if E. coli and Buchnera showed the same order and orientation apart from missing genes in Buchnera, and if these missing genes could not be located elsewhere in the Buchnera genome (example in Figure 2). Syntenic regions always terminated with loci that were present in both species.

This criterion resulted in 91 fragments in the ancestor that contained at least two genes and that corresponded to regions of synteny between Buchnera and E. coli (Figures 3,4). In addition, there were 62 single genes. The longest syntenic fragment, containing the megaoperon of ribosomal proteins that is widely conserved among bacteria, consisted of 89 kb spanning 77 genes in the ancestor and 45 genes in Buchnera. In addition, there were 143 ancestral regions between these retained fragments, containing genes inferred to be lost in Buchnera. (This number is slightly less than the number of retained fragments (153) because, whereas most of the retained fragments were flanked by lost regions, a few directly bordered other retained fragments.) The ancestral genome was therefore divided into a total of 306 fragments, just over half of which were retained in the Buchnera genome, and there are 306 junctions between these fragments on the circular chromosome (Figure 3).

The evolution of Buchnera was clearly accompanied by many chromosome rearrangements and massive changes in genome size (Figures 3,5). Remarkably, a comparison of ortholog positions between Buchnera and E. coli indicates that a pattern of approximate symmetry of gene distance from the replication origin and terminus was maintained (Figure 6). This is indicated by an X-pattern when the positions in the two genomes are plotted against one another with the origins of replication as endpoints (the replication origin in Buchnera was designated by Shigenobu et al. [8] on the basis of the position of the only DnaA box present on the chromosome and on a shift in the GC-skew value at third codon positions around that region). This X-pattern has recently been found to be typical for comparisons of orthologous regions between pairs of closely related bacteria [18,19,20]. It is most readily explained as the result of successive inversions around the replication terminus or origin.

On the basis of the reconstructed ancestor, a total of 503 genes in 156 locations were lost in deletions within regions of synteny (Figure 7, top). These deleted regions are sometimes large, with 11 regions spanning 10 or more genes (Figure 7, top). A much larger proportion of the reduction can be attributed to regions lost between syntenic fragments; 1,449 genes were lost in 143 such gaps (Figure 7, bottom).

For losses occurring both within and between syntenic fragments, a single lost region frequently contains multiple genes, often involved in seemingly unrelated functions (see, for example, Figures 2 and 5).

Some genes deleted within regions of synteny persist as partially degraded sequences that range from pseudogenes having clear homology with the E. coli ortholog [8], to much shortened sequences with no recognizable sequence homology. These sequences indicate that function has sometimes been eliminated, initially through a small deletion or substitution, with the remaining DNA sequence subsequently eroded by multiple mutations that are biased in favor of deletion over insertion, as documented for Rickettsia [7,9] and other bacteria including Buchnera [21]. An analysis of Buchnera pseudogenes showed that deletions outnumber insertions (31 versus 2) and that the number of nucleotides deleted is over 100-fold greater than those inserted [21].

An expected result of this gradual gene degradation is that intergenic spacers that contain gene remnants will be longer than spacers where no gene loss has occurred. Buchnera spacers can be categorized into three groups: those flanked by the same genes in Buchnera and the ancestor, those that occur within regions of synteny at positions where gene(s) are missing in Buchnera, and those that occur between syntenic fragments. Spacers in the first category, which can be considered to be descended from ancestral spacers, have an average length of 55 bp (n = 272). These ancient spacers are much shorter than spacers occurring where ancestral genes have been deleted within syntenic fragments (188 bp, n = 165) or than spacers occurring between syntenic fragments (also 188 bp, n = 162). Considering only spacers within syntenic fragments, spacer length does not increase further when the number of genes lost is greater than one (Figure 8).

Small bacterial genomes typically contain a smaller proportion of regulatory elements than do larger genomes [22,23]. Consistent with this trend, the genome of Buchnera has been noted to lack virtually all regulatory proteins [8]. We examined promoters within intergenic spacers flanked by the same genes in both Buchnera and E. coli under the assumption that these evolved from a common ancestral sequence and thus represent orthologous spacers. In 44 of the orthologous spacers, a σ-70 promoter is annotated in E. coli on the basis of experimental evidence and/or presence of a promoter consensus sequence [14]. The consensus sequence for this promoter is TTGACA-(l7 nucleotides)-TATAAT, from which the -35 and -10 regions are each defined primarily by three base pairs:TTG--- for the -35 region and TA---T for the -10 region [24,25]. We applied a conservative criterion for presence of a promoter in Buchnera, by requiring retention of as few as four of these six nucleotides with a separation of 16-18 nucleotides. In 16 of the 33 spacers in which flanking genes are encoded on the same strand and in which E. coli has an annotated promoter, Buchnera lacks any recognizable promoter. All such apparent losses occurred when the flanking genes were encoded on the same strand, suggesting that promoter loss is frequent when a group of newly contiguous genes can be translated as a single transcriptional unit. In some cases, fusion of genes into the same polycistron has occurred following the loss of intervening genes oriented in the opposite direction (Figure 9). In contrast, in none of the 11 cases in which orthologous spacers were located between genes oriented in opposite directions did Buchnera lose promoters.

On the basis of sequence criteria, Buchnera also shows degeneration of the Shine-Dalgarno (SD) sequences that promote initiation of translation by binding with the anti-SD complement in rRNA (which is widely conserved and identical for Buchnera and E. coli). Changes in the SD sequence can impede initiation of translation [26,27]. In comparisons of spacers orthologous between E. coli and Buchnera, Buchnera showed a lower number of matches to the core eight nucleotides of the SD sequence in 37 of 51 cases (73%). Also, a total of 20 protein-binding sites were annotated within E. coli spacers orthologous to Buchnera spacers (that is, with the same flanking genes in the same orientation). In Buchnera, only one was preserved, four had half of the sequence maintained, and the remaining 15 were not detectable.

These hypothetical losses of regulatory regions from the reduced genome of Buchnera are based on levels of sequence similarity with the known consensus sequence in modern E. coli. There is no experimental evidence to eliminate the possibility that the degraded promoters in Buchnera are still functional, or that the symbiont uses different promoters from those identified in E. coli. In addition, inaccuracies in the annotated positions of Buchnera genes, in which positions of start codons are not verified experimentally, could bias analyses of SD sequences.

Endosymbiosis or chronic pathogenesis involves a metabolite-rich environment within host tissues, low growth rates and isolation of strains within host individuals. This combination of factors results in relaxed selection at many loci and higher levels of genetic drift affecting the entire genome. The reduced effectiveness of selection is reflected in patterns of sequence evolution in Buchnera and other endosymbionts [17,28,29]. It is also expected to affect genome evolution, accelerating the loss of genes that are entirely superfluous as well as those that are beneficial but not essential.

Small genome size itself is likely to reflect lack of selection for gene retention rather than direct selection for a compact genome [21]. Especially in the case of Buchnera, which possesses 50-200 chromosomes per cell [30] and in which nonfunctional pseudogenes can persist for long periods [31], selection for reduced DNA content seems an untenable explanation for its small genome. The polyploidy itself may be a consequence of the loss, through genetic drift, of one or more loci involved in regulating and resolving chromosome replication during the cell cycle.

One pattern that has emerged from comparative analyses of fully sequenced genomes is that, although different pathogen and symbiont lineages approach the same minimal genome size, the inventories of genes retained by these organisms are extremely different [3,32,33,34,35,36]. Relatively few genes are universally distributed, and universal cellular processes depend in part on nonorthologous genes. The divergence in gene inventories among small-genome bacteria implies redundancy in ancestral genomes underlying central cell processes, including DNA processing, transcription and translation. For the most part, this redundancy does not arise from the presence of paralogous genes, which are relatively few in bacteria. An examination of the process of genome reduction, as exemplified by Buchnera, could yield insight into why gene inventories differ among small genomes.

The comparison of the Buchnera genome to the reconstructed ancestral genome suggests that a considerable part of genome reduction occurred through large deletions accompanying chromosome rearrangements. One indicator that supports gene loss partly through large deletions spanning multiple genes is the distribution of sizes of these regions in the ancestor (Figure 7, bottom). If genes were lost one by one, the expectation is that retained genes would be randomly mixed with lost genes within the ancestral genome. In the observed distribution, lost genes are aggregated. Although part of this clustering might be attributed to the linkage of functionally related genes, this would not account for the larger segments. Another indicator of gene loss through large deletions is the lack of positive relationship between spacer length and number of genes lost for gene deletions that occurred within syntenic fragments (Figure 8). Thus, the most plausible explanation for much of the gene loss occurring in the evolution of Buchnera is the fixation of single large deletions spanning many genes, although it is also clear that some genes were eliminated through smaller deletions and gradual erosion. Large deletions spanning multiple genes are possible only early in the reductive process, when many genes are nonessential. The content of initial large deletions will determine the degree of selection on remaining loci and thus will govern the ultimate composition of the reduced genome.

One distinctive aspect of the Buchnera genome is the presence of only one copy each of the genes encoding 16S, 23S and 5S rRNA, and the separation of these genes into two transcriptional units with the 16S rRNA gene (rrs) apart from the others (rrf and rrl) [37]. Typically, bacteria possess multiple rRNA operons, each containing all three genes; the Buchnera ancestor is inferred to have possessed at least five operons and modern E. coli has seven. The remaining rRNA genes descend from two different ancestral operons, with other rRNA operons lost entirely. From flanking genes, it is evident that the two retained transcription units correspond to rrsH and rrfD-rrlD of E. coli. The missing parts of these operons (rrfH, rrlH, and rrsD) were lost in deletions within syntenic fragments. All other rRNA operons were lost in entirety as part of regions occurring between syntenic fragments. Thus, the modern number and arrangement of rRNA genes is the result of the loss of large regions, possibly in the course of rearrangements, as well as the elimination of individual genes within operons.

E. coli has 86 tRNAs, and most of these appear to be present in the ancestor (the parsimony criterion for presence of a tRNA in the ancestor, based on distribution in sequenced genomes, is somewhat unreliable because of sequence homology among different tRNA genes). Buchnera has only 32 tRNAs, with most amino acids having only one. Assuming that the ancestor had the same complement of tRNAs as E. coli, 15 were lost from within syntenic fragments and 39 were lost in regions between syntenic fragments. Selection enforces the retention of at least one tRNA for each amino acid, as observed in modern Buchnera, so the early fixation of large deletions containing some tRNAs would have created a selective requirement for the retention of others.

One example of a functional category that is routinely reduced in small-genome bacteria is that consisting of genes for DNA repair and recombination. As is typical for small genome bacteria, Buchnera has lost a large number of repair genes (Table 1). In this functional category, as in others, the set of retained genes differs among reduced genomes [15]. For example, Buchnera is unique among sequenced genomes in having lost recA, which functions in homologous recombination and repair. In contrast, Buchnera retains recBCD, which has been lost by several other small genomes [8,12,15]. The reconstructed events underlying the loss of individual repair genes include many large deletions, encompassing many genes (Table 1). For example, under the reconstruction, recA is part of a contiguous deleted region of about 10 kb containing ten genes. Also unusual in Buchnera is the lack of uvrA, uvrB and uvrC (encoding an excision nuclease involved in repair of UV damage to DNA). The uvrB and uvrC genes fall in separate large deleted regions between syntenic fragments, whereas uvrA is one of six genes deleted within a syntenic region, with the corresponding position in Buchnera occupied by a 618 bp spacer (Table 1).

The most plausible interpretation of this pattern is that some of these repair functions were initially lost as the result of large deletions, sometimes occurring in combination with chromosomal rearrangements. These deletions were followed by gradual loss of other genes in the same pathways. The process of fixation of these large deletions reflects not only selection on the repair functions but also selection on other genes on the deleted fragments. A gene flanked by required loci is less likely to be lost in an early large deletion than a gene flanked by nonessential loci. For example, if the 10 kb segment containing recA was lost as a single deletion, then the loss of recA was dependent on the fact that the neighboring genes included in the deletion were not essential for survival. The retention of recBCD might have been promoted by Buchnera's requirement for the flanking gene, argA, which encodes an enzyme for biosynthesis of arginine. Other small-genome bacteria lack argA and other genes underlying amino-acid biosynthesis, whereas Buchnera retains all genes required for biosynthesis of essential amino acids, which are needed by the host.

Many genes were lost singly, implying insufficient selection for conservation of individual genes. A consequence is the correlation in presence/absence among genes in the same pathway, even if they occupy separate locations on the chromosome: if pathway function is lost, all genes in the pathway are invariably lost or degraded. An example in the case of the recombinase genes is the loss of recF, which is required for some recA functions [38] and which was deleted individually with both flanking genes retained (Table 1). A plausible scenario for the loss of this pathway is that, once recA was eliminated in the context of a large deletion, there was no selection for recF retention and it was subject to successive small deletions and substitutions. In both Buchnera of A. pisum and Buchnera of Schizaphis graminum, recF is replaced by 127-138 nucleotides, with base composition 87-89% A+T [8,39]. The process of shrinkage and A+T accumulation could result strictly from mutational patterns, which are biased towards deletions [21] and nucleotide substitutions ending in A or T [31].

Genome comparisons across other groups of related bacteria suggest that deletion events encompassing multiple loci are frequent in bacterial evolution. For example, deletions of over 20 kb and 20 loci have occurred in natural isolates of E. coli that are very closely related on the basis of sequence homology [40]. On the basis of comparison with its close relative Mycobacterium tuberculosis, M. leprae shows a pattern of genome reduction that includes both loss of large fragments and also loss of gene function through slight changes in sequence [41]. Different strains of M. tuberculosis have been found to contain at least 25 long deletions, with one event removing as many as 16 ORFs [42].

The changes in sequences underlying initiation of both transcription and translation, as well as the loss of regulatory proteins [8], give an impression of genome-wide degeneration of regulatory functions in Buchnera. The hypothesized elimination of promoters and genes seems to have produced newly formed polycistronic regions (Figure 9). The fusion of genes into single transcriptional units has been interpreted as the result of selection favoring small genome size [43] or favoring efficiency in transcription or translation [44]. In Buchnera, however, the hypothesized loss of promoters suggests a general genomic decay influencing transcriptional regulation; this is supported by the observation that SD sequences are also degenerate, and by the finding that some of the most frequently used codons in the genome do not have corresponding tRNAs because they have been deleted throughout Buchnera's evolution. Experiments on gene expression patterns are needed to test the potential deterioration of regulatory capabilities.

Phylogenetic studies based on 16S rDNA sequences indicate that Buchnera belongs to the gamma-3 Proteobacteria and is closely related to E. coli (Figure 1). The precise phylogenetic position of Buchnera has varied depending on the method of analysis and the taxa included in the analysis. This instability is partially the result of the accelerated evolution and AT bias that affect all Buchnera genes including the 16S rDNA [17]. Most studies place Buchnera either within or just outside the Enterobacteriaceae, suggesting that it diverged near the time of the common ancestor of this clade [48,49,50,51]. Thus, the reconstructed ancestor of the clade corresponding to Enterobacteriaceae (A in Figure 1) was considered to constitute a close approximation of the free-living ancestor that gave rise to Buchnera.

The genes absent from Buchnera and present in E. coli include many that are present in other related bacteria such as Haemophilus influenzae and V. cholerae. This supports the view that Buchnera was derived through loss of genes from an ancestor similar to E. coli, a conclusion also reached by Shigenobu et al. [8]. However, some genes were acquired recently by E. coli, after divergence from the Buchnera lineage [52]. The ancestral genome reconstruction was based on the E. coli MCl655 genome [14], following the removal of genes that seemed to have been acquired after the E. coli-Buchnera divergence. The rules for removing a gene were based on phylogenetic distribution of orthologs among the unannotated genomes of three serovars of Salmonella enterica (sv. Typhi, sv. Paratyphi and sv. Typhimurium), Klebsiella pneumoniae and Yersinia pestis and the annotated genome of V. cholerae (GenBank accession number NC 002505) [53]. The unpublished sequence data were produced by the Salmonella, Klebsiella and Yersinia Sequencing Groups at the Sanger Centre (for S. enterica sv. Typhi, K. pneumoniae and Y. pestis) [54] and the Washington University Genome Sequencing Center (for S. enterica sv. Typhimurium and S. enterica sv. Paratyphi) [55]. Determination of phylogenetic distribution was based on the 'Pip-maker' analyses [56] for viewing orthology among enteric and related bacteria as produced by McClelland et al. [57,58]. To qualify as an ortholog to an E. coli gene, a sequence had to show at least 40% amino-acid identity for at least half of the E. coli gene.

Inclusion of a gene in the ancestor was not dependent on its presence in all the descendant taxa, as a gene could be ancestral but lost from some lineages or acquired after the ancestor but before divergence of some descendant taxa. Therefore, to be retained in the ancestor, an ortholog had to be present in at least one of the close relatives of E. coli (S. enterica sv. Paratyphi, Typhi or Typhimurium, or K. pneumoniae) and also in at least one of the more distant relatives, either Y. pestis (representing a more basal branch of the Enterobacteriaceae) or V. cholerae (branching just outside the Enterobacteriaceae). These requirements effectively impose a parsimony criterion, minimizing the number of gains and losses of genes. This criterion is likely to be valid for most genes across this relatively shallow phylogeny. Sequences most likely to be misrepresented in the ancestor by this reconstruction are phage genes or insertion elements, which may move in and out of genomes too frequently for ancestral states to be reconstructed using parsimony. Buchnera lacks phage or insertion sequences [8].

The ancestor was assigned the gene order of modern E. coli. Although this order is almost certainly not entirely correct, comparison with other bacteria indicates that it is true for a majority of the borders between the 306 fragments (including those lost and those retained). For example, the same linkages occur in E. coli and V. cholerae and/or E. coli and Y. pestis for 219 of the junctions between fragments in Buchnera (example in Figure 5). This implies that most of the E. coli arrangements were present in the ancestor of Buchnera and E. coli and that the ends of syntenic fragments mostly represent sites at which rearrangements occurred in the evolution of Buchnera. A concentration of rearrangements in the Buchnera lineage is also suggested by the much longer average length of spacers that occur between syntenic fragments as compared to spacers that have the same flanking genes in the ancestor and in E. coli (see above). This longer spacer length can be interpreted as arising from remnants of genes that were rendered functionless when rearrangements occurred. When the relevant genomes currently in progress are complete and annotated, a more accurate reconstruction of ancestral gene order maybe possible.