Inferring Genome Trees by Using a Filter To Eliminate Phylogenetically Discordant Sequences and a Distance Matrix Based on Mean Normalized BLASTP Scores

Darwin's paradigm holds that the diversity of present-day organismshas arisen via a process of genetic descent with modification,as on a bifurcating tree. Evidence is accumulating that genesare sometimes transferred not along lineages but rather acrosslineages. To the extent that this is so, Darwin's paradigm canapply only imperfectly to genomes, potentially complicatingor perhaps undermining attempts to reconstruct historical relationshipsamong genomes (i.e., a genome tree). Whether most genes in agenome have arisen via treelike (vertical) descent or by lateraltransfer across lineages can be tested if enough complete genomesequences are used. We define a phylogenetically discordantsequence (PDS) as an open reading frame (ORF) that exhibitspatterns of similarity relationships statistically distinguishablefrom those of most other ORFs in the same genome. PDSs representbetween 6.0 and 16.8% (mean, 10.8%) of the analyzable ORFs inthe genomes of 28 bacteria, eight archaea, and one eukaryote(Saccharomyces cerevisiae). In this study we developed and assesseda distance-based approach, based on mean pairwise sequence similarity,for generating genome trees. Exclusion of PDSs improved bootstrapsupport for basal nodes but altered few topological features,indicating that there is little systematic bias among PDSs.Many but not all features of the genome tree from which PDSswere excluded are consistent with the 16S rRNA tree.

INTRODUCTION

Top

Introduction

Since the early days of molecular biology, it has been assumedthat the historical pattern of genetic relationships among organismscould eventually be reconstructed via statistically based comparisonsof the sequences of genes or proteins within individual families(47). In keeping with Darwin's paradigm of treelike descentwith modification, these relationships are represented as bifurcatingtrees. If a gene or protein family has not been affected bygene duplication (i.e., if only orthologs are being compared),its phylogenetic tree shows organismal relationships as well.With the recent rise of genomics, attention has been focusedon the genome as an entity that is conceptually distinguishablefrom both genes and organism (18). As most genes are physicallycolocated on chromosomes that partition together during meiosisand mitosis, however, the default assumption has been that genetic,genomic, and organismal phylogenies are merely different facetsor delineations of the same fundamental historical process.The information content of individual genes may sometimes beinsufficient to resolve the complete details of this process,but sequences of multiple genes or proteins—now availablefrom complete genome sequences—can be analyzed jointlyto obtain a better (historically more accurate) tree.

However, a potentially serious complication has arisen. It isnow appreciated that some genes are sometimes transmitted notvertically (along individual branches of the organismal treethrough time) but laterally or horizontally (directly from onebranch of the tree to another). The closest relatives (nearestorthologs) of a laterally transferred gene thus occur in genomesin the organismal lineage from which the transfer originated,not in the relatives of its new host; as a consequence, theextrapolated organismal phylogeny is incongruent with the phylogenyinferred from families whose members have been transmitted onlyvertically. If such lateral gene transfer (LGT) were frequent,the number and diversity of anomalous gene trees might erodeour ability to reconstruct correct organismal trees or perhapseven empty this concept of meaning. In the extreme case, whatwe consider species might have been generated and be maintainednot by common ancestry but rather by barriers and firewallsto genetic recombination 38; W. F. Doolittle, personal communication).Many studies have suggested that LGT has occurred frequentlyin many prokaryotic lineages (7-9, 16, 19-21, 24-26, 31, 33,38).

The degree to which our understanding of genome evolution maybe compromised by LGT depends on the extent to which geneticmaterial has been transferred laterally and, more generally,on our ability to detect and control for incongruent data. Wehypothesize that removal of incongruent data should leave setsof genes which, analyzed jointly, should yield a phylogeny thatbetter corresponds with organismal (phenotypic, ultrastructural,and physiological) groupings and with trees inferred from genefamilies that have undergone little or no lateral transfer (e.g.,small-subunit rRNA genes [43]). If the incongruent data aretopologically biased (e.g., if they can be identified with oneor a few major LGT events), eliminating them from the analysismight yield a topologically different tree. On the other hand,if the incongruent data are unbiased (e.g., if they arose froma large number of dissimilar, quantitatively less-significantevents), their removal might improve the statistical supportfor subtrees but have little effect on tree topology. Thesecompeting predictions can be tested directly by inferring whole-genometrees before and after suspect data have been identified andeliminated.

Phylogenetic trees are inferred for individual gene or proteinfamilies by maximizing or minimizing an appropriate functionover all putatively homologous positions. Trees of organismsare typically derived by parsimony analysis of discrete characterstates. The first genome trees were constructed by distanceanalysis and were based on distances derived from proportionsof statistically similar open reading frames (ORFs) (presumptiveorthologs) or protein folds that could be recognized by pairwisecomparisons of genomes (14, 27, 41, 42, 44). Basing these treeson shared ORF or fold contents avoided the perhaps intractablecomplexities of sequence-based alignment of entire genomes butfailed to capture information about the degrees to which setsof ORFs (presumed orthologs) are similar. Grishin et al. (17)used a distance measure based on empirically determined distributionsof pairwise interprotein amino acid substitution rates for genomes,while Wolf et al. (44) examined transformations of the pairwisepercentages of identity between orthologs. Here we utilizedan alternative method of constructing genome trees based insteadon mean normalized BLASTP (1) scores, and we compared treesproduced by this approach with a content-based genome tree.This study did not constitute a test of the idea that lineagesare artifacts of recombinational barriers rather than of shareddescent but did explore the extent to which sequences that areincongruent for any reason with the majority signals from theirown genomes erode support for a single common phylogenetic tree.

Lateral genetic transfer is not the only process that can obfuscatea phylogeny. An ORF can be phylogenetically discordant (i) ifits orthologs have been lost in some but not all other genomes,leaving a patchwork of orthologous and paralogous matches; (ii)because of convergent evolution and nonneutral evolution ingeneral; and (iii) in cases in which certain genes exhibit ratesor patterns of sequence change substantially different fromthose of the other genes in the lineage. To avoid circularityin using phylogenetic analysis to assess phylogenetic incongruity,in this study we instead developed a pairwise statistical approach,correlating patterns of observed genomic similarity among species.

MATERIALS AND METHODS

Top

Materials and Methods

We established the target database for the initial BLASTP (1)analysis by parsing the GenBank database by species to obtaina large number of "virtual genomes." Each query database (onefor each genome in our analysis) consisted of all of the protein-encodingORFs recognized for the genome by the respective genome projector community. Each query ORF found a best BLASTP match betterthan a defined threshold in none, one, or more target speciesamong the species for which there are data in the GenBank database.We ranked these best matches by strength (i.e., improbabilityof occurrence by chance). Phylogenetic discordance among ORFsin a query genome or between a query ORF and the bulk signalfrom its own genome was assessed by statistical comparison ofthe rankings (see below). As explained above, it is the patternof orthologs that marks an ORF as having arisen by LGT. However,orthology is a tree-based concept (13) that usually cannot bedirectly established by pairwise comparison. Following establishedpractice (30, 41, 44), we focused instead on reciprocal bestmatches (RBMs). If a query ORF has no ortholog in a given targetgenome, BLASTP analysis might nonetheless identify a best matchfor it (probably a paralog), but the paralog's best match backinto the query genome would probably be its own ortholog andnot the original query ORF (i.e., the match would be nonreciprocal).This approach is imperfect, as it can be confounded by the absenceof a target ORF's true ortholog in the query genome and bestBLAST matches do not need to be phylogenetically adjacent sequences(22); nonetheless, it should be unbiased and eliminate manyinstances of paralogy. Query ORFs involved in fewer than fourRBMs were excluded from further analysis, as such patterns containedtoo little information for rigorous comparison (see below).

If all genes within a genome have the same phylogenetic history,the RBMs for each gene vis-a-vis the target genomes (here, theGenBank virtual genomes) should rank similarly. The rankingsshould be the same, within statistical variation, for the genomeas a whole and for each constituent ORF. If an ORF does nothave this common ranking but instead shows a conflicting pattern,it is discordant. Indeed, it is phylogenetically discordantbecause, by analogy with construction of a tree from a distancematrix, a conflicting pattern of similarity relationships mustspecify a conflicting tree. Missing orthologs create gaps butnot misinformation (incorrectly ordered rankings).

To quantify these relationships, we introduced a normalizedsimilarity score (u) for a query ORF and its RBM in a giventarget species. We computed this score by dividing the BLASTP-basedsimilarity score (bit score [S']) by the ORF's self-matchingscore (Fig. 1). The median of the u values for a target speciesdefined w, a measure of the query genome's overall sequencesimilarity with sequences from that target species. By correlatingu and w for all species in which the ORF found a match at aBLASTP expectation value greater than a defined threshold (heree = 1.0 x 10^-10), we could determine if the pattern of relationshipswas consistent with the ORF having evolved concordantly withthe rest of the genome.

View larger version (21K):
[in this window]
[in a new window]
FIG. 1. Statistical strategy. In this schematic diagram, ORF 1 finds RBMs in species 1, 2, and S; ORF 2 finds RBMs in species 1, 3, and S; and ORF L finds RBMs in species 2 and S. Thus, w₁ = {u_1,1, u_2,1, . . .}; w₂ = {u_1,2, . . ., u_L,2}; w₃ = {u_2,3, . . .}; and w_S = {u_1,S, u_2,S, . . ., u_L,S}. In this example, ORF 1's set of w values is {w₁, w₂, . . ., w_S}; ORF 2's set of w values is {w₁, w₃, . . ., w_S}; and ORF L's set of w values is {w₂, . . ., w_S}. An ORF's set of u values is correlated with its set of w values as described in the text.

The set of w values associated with a given ORF formed the intrinsichypothesis of the test; that is, it was the relative distributionof w values against which we tested the distribution of u values.Since the distributions of u values from which each w was calculatedwere not identical, we used an approximate randomization test(32) to determine the significance of the observed correlationbetween u and w. We chose the Spearman rank correlation (37,46) as our test statistic because there was no reason to assumethat the relationship between u and w was linear. To approximatethe distribution of the test statistic, the underlying distributionof u values that made up each w was sampled with replacementto generate random sets of u values that were then correlatedwith w. The probability of obtaining a correlation (r) valueas small as or smaller than the observed r₀ value was definedby (C + 1)/(N_R + 1), where C is the count of r_R $<=$ r₀ and N_R isthe total number of randomizations. (N_R was the lesser of 199,999and ${Pi}$ n_i, ${forall}$ n_i > 0, where n_i is the number of ORFs shared bythe query genome and the target genome [i].) For each pair ofgenomes (the query and each target), the size of each resampledset was equal to the number of RBMs (Table 1).

View this table:
[in this window]
[in a new window]
TABLE 1. Proportions of ORFs analyzable and proportions of ORFs found to be phylogenetically discordant in each genome

The 16S rRNA tree was extracted from a much larger (7,322-prokaryoticsequence) maximum-likelihood tree maintained by the RibosomalDatabase Project (29).

Genome trees were inferred by Fitch-Margoliash least-squaresanalysis of distance-type matrices (17), each element of whichwas an ORF-based measure of pairwise dissimilarity between genomes,either 1.0 minus the proportion of ORFs shared by a given pairof genomes (41) or 1.0 minus the mean of normalized pairwiseBLASTP scores. BLASTP scores (42) were normalized (4) by dividingan ORF's score against the target genome by the ORF's scoreagainst itself. The target database was the set of ORFs identifiedfor the genome itself, not, as described above, all genes depositedunder the species designation in GenBank. Only ORFs with matchesbetter than a defined threshold (BLASTP e = 1.0 x 10^-10) wereused; the reciprocal best-match criterion was not used to derivegenome trees. Distance matrices were generated using the NeuroGadgetsInc. web service (http://www.neurogadgets.com). Distance analysiswas carried out using the FITCH program in the PHYLIP softwarepackage (11), with global rearrangements and randomized (jumbled)species input order. For bootstrap analysis, samples (n = 100)were taken with replacement from the ORFs shared by each pairof genomes, and from these a mean distance was calculated forthat pair of genomes. Distance trees were again generated usingFITCH, and the majority rule consensus tree was computed usingthe CONSENSE program in PHYLIP. Trees were visualized and bootstrapvalues were added using TREEVIEW (36).

RESULTS

Top

Results

A total of 81,160 ORFs in 37 query genomes representing theBacteria (28 species or strains), the Archaea (eight species),and Saccharomyces cerevisiae were examined; for 42,925 (52.9%)of these ORFs we found four or more RBMs, and we used theseORFs for analysis (Table 1). The minimum analyzable proportionwas found in Aeropyrum pernix (31%), and the maximum analyzableproportion was found in Buchnera sp. strain APS (94%); the unweightedmean for 37 genomes was 56%. Of the 42,925 analyzable ORFs,4,331 (10.1%) were phylogenetically discordant at P < 0.05.The lowest proportion of phylogenetically discordant sequences(PDSs) was observed in Pyrococcus horikoshi (6.0%), and thehighest proportion of PDSs was observed in Treponema pallidum(16.8%); the unweighted mean for these genomes was 10.8%. Wesuspect that many of the sparsely distributed ORFs excludedfrom this analysis are phylogenetically discordant as well (39).

As described above, the first genome phylogenies were basedon a distance-type measure derived from the proportions of ORFsshared pairwise by genomes. Figure 2 shows a genome tree forthe 37 microbial genomes based on the proportion in each querygenome of ORFs that had an initial BLASTP match better thanthe threshold (e = 1.0 x 10^-10) in each other genome. It wasnot immediately obvious how to assess statistical support (confidenceintervals for internal nodes) for such a tree. Snel et al. (41)assessed confidence by the half-delete jackknife method (45),deleting random halves of ORFs in each query genome, reassessingthe proportions of shared genes, inferring a new tree for eachreplicate, and counting the number of times out of 100 thata particular cluster was found. Wolf et al. (44) instead usedthe nonparametric bootstrap method (12), resampling from amongidentified orthologs. The proportion-based genome tree for the37 genomes (Fig. 2) resembles the tree based on 16S rRNA sequences(Fig. 3) in many respects, but there are a number of discrepancies.In the proportion-based genome tree the Thermotoga maritimagenome branches basally among bacterial genomes, as the Thermotoga16S rRNA does. However, in the proportion-based tree, the genomeof Aquifex aeolicus does not branch second deepest; instead,it groups with the genome of Synechocystis as the third-deepestbranch. The Mycobacterium tuberculosis genome does not groupwith the genomes of other Firmicutes; instead, it joins thegenome of Deinococcus radiodurans on the second-most-basal branch.In addition, the Haemophilus influenzae genome does not groupwith the genomes of enteric members of the ${gamma}$ subclass of theclass Proteobacteria ( ${gamma}$ -proteobacteria). Even more discrepantis the archaeal subtree, in which the genomes of Halobacteriumsalinarum and A. pernix constitute the two most-basal branches(the former does not group with genomes of other euryarchaeotes).The methanogen genomes, together with the genomes of Archaeoglobusfulgidus and the two Pyrococcus species, form a group not seenin 16S rRNA trees or indeed in most other molecular sequencetrees.

View larger version (33K):
[in this window]
[in a new window]
FIG. 2. Genome tree based on the proportions of query ORFs that found a match in each target genome better than the threshold BLASTP expectation (e = 1.0 x 10^-10). Proportions were computed (http://www.neurogadgets.com) by determining the number of ORFs shared by the query genome (smaller of the pair) and the target genome (larger of the pair) and then dividing this number by the number of ORFs in the query genome. A distance matrix was generated by computing 1.00 minus the proportion of shared loci; this matrix was imported into PHYLIP (12) for analysis by FITCH (see Materials and Methods). The branch lengths reflect distances, as assessed by these criteria, between genomes.

View larger version (25K):
[in this window]
[in a new window]
FIG. 3. 16S rRNA gene tree, adapted from the Ribosomal Database Project (29). Where 16S rRNA sequences were not available from the Ribosomal Database Project, close relatives were selected. The major organism classifications are consistent with those described by Olsen et al. (34).

To explore whether the apparent anomalies arose because theunderlying similarity matrix was based on proportions of sharedORFs (instead of some other measure of distance), we developedan alternative measure based on mean sequence similarity (seeMaterials and Methods). Again, only the ORFs having a matchbetter than the threshold (BLASTP e = 1.0 x 10^-10) were considered.We assessed confidence in internal nodes by using the nonparametricbootstrap method (11, 44), sampling with replacement from theORFs shared by each pair of genomes to generate the mean distancesused to construct trees. This approach lends itself to filtering(see below) better than the approach based on proportions ofshared ORFs does. The genome tree produced by using this alternativemethod (Fig. 4) differs from the proportion-based tree (Fig.2) but also exhibits some different discrepancies vis-a-visthe 16S rRNA tree (Fig. 3).

View larger version (39K):
[in this window]
[in a new window]
FIG. 4. Genome tree based on normalized BLASTP scores, constructed by using all query ORFs that found a match in the target genome better than the BLASTP expectation (e = 1.0 x 10^-10). The tree was constructed as described in the legend to Fig. 2, but 100 random replicates (pairwise genome scores reconstructed from individual ORF scores, with resampling) were examined. The numbers at the nodes are bootstrap values, which were generated by the CONSENSE program in PHYLIP.

In the genome tree based on mean normalized BLASTP scores (Fig.4), the T. maritima genome retains its position as the most-basalbranch among the bacterial genomes; the A. aeolicus genome is(as it is in the 16S rRNA tree) resolved with good confidenceas the second-deepest branch; and the genome of D. radioduransappears third. Chlamydial genomes form the sister group of spirochetegenomes as they do in the 16S rRNA sequence tree, although cyanobacterialgenomes are still distant. A large clade of genomes from Proteobacteria,low-G+C-content Firmicutes, and chlamydiae plus spirochetesis well supported, and the genome of the high-G+C-content firmicuteM. tuberculosis is completely outside the group. The genomeof Xylella fastidiosa is at the base of the ${gamma}$ -proteobacterialgenome group, as it is in the proportion tree, although thisgroup includes genomes of ß-proteobacteria (representedby the two Neisseria meningitidis genomes). 16S rRNA sequencetrees (Fig. 3) (34) also suggest that the ß-proteobacterialsequences might be included among the sequences of the ${gamma}$ -proteobacteria.The genome of H. influenzae groups with the genomes of the otherenteric ${gamma}$ -proteobacteria, as it does in the 16S rRNA tree, whichis consistent with the findings of nonmolecular bacterial systematics.The archaeal portion of the score-based genome tree is rearrangedcompared to the archaeal portion of the proportion-based genometree (Fig. 2) and is topologically quite different from thearchaeal portion of the 16S rRNA tree (Fig. 3), with genomesof both H. salinarum and Thermoplasma acidophilum branchingearlier than the genome of the crenarchaeote A. pernix.

Excluding sequences that could not be analyzed for potentialphylogenetic discordance (but retaining the discordant sequences)yielded a tree that was essentially identical to the tree shownin Fig. 4 (data not shown) and differed in only three topologicalfeatures. The group of five Bacillus, Mycoplasma, and Ureaplasmagenomes was resolved as a sister group of the spirochete andchlamydial genomes, albeit with very weak bootstrap support(55%), instead of branching next from the backbone as in Fig.4. The genome of Pseudomonas aeruginosa branches immediatelyafter the two Neisseria genomes, instead of forming a sistergroup with them as in Fig. 4, although the bootstrap supportis weak in both cases (67 and 62%). Finally, the branching orderof the H. influenzae and Buchnera sp. genomes is reversed, againwith low bootstrap support. Thus, excluding the 47.1% of theORFs that had relatively few strong matches in these genomeshad an almost negligible effect on the topology of the genometree based on mean normalized BLAST scores.

Exclusion of PDSs (P < 0.05) from the latter set yieldedthe tree shown in Fig. 5. The topology of this tree is almostidentical to that of the tree described above, differing onlyin the position of the P. aeruginosa genome, which groups weaklywith the two Neisseria genomes, as in Fig. 4. The two othertopological changes compared with Fig. 4 resulted from excludingthe 38,235 nonanalyzable ORFs, not from removing the 4,331 discordantORFs. Exclusion of discordant sequences did, however, substantiallyimprove the resolution of some subtrees, as measured by thenonparametric bootstrap method. Within the ${gamma}$ and ß-proteobacteria,for example, the mean bootstrap values increased from 86.1%(Fig. 4) and 86.6% (in the tree containing all analyzable ORFs[data not shown]) to 91.3% (Fig. 5).

View larger version (36K):
[in this window]
[in a new window]
FIG. 5. Genome tree based on normalized BLASTP scores better than the threshold (e = 1.0 x 10^-10), after removal of PDSs (P < 0.05). Tree construction and bootstrapping were performed as described in the legend to Fig. 4.

DISCUSSION

Top

Discussion

Our results demonstrate that a biologically reasonable, statisticallywell-supported genome tree can be produced from a distance typeof matrix based on mean normalized BLASTP scores. BLAST (1)is based on pairwise local alignments and consequently doesnot guarantee that all homologous characters are recognizedand scored as such; on the other hand, normalized BLAST similarity(bit) scores are well defined statistically and can be comparedand ranked much more naturally than global alignments can. Byanalogy with algorithms by which trees are derived from distancematrices, we hypothesized that it is possible to define statisticallya difference in rankings among elements within individual rowsor columns of distance (here, BLASTP-based) matrices such thatmatrices with anomalously ranked elements would almost alwaysspecify an incongruent tree. Our results indicate that despiteits inherent approximations, our approach was successful enoughto warrant further development and refinement (e.g., iterativecomparisons and use of a rescaled distance measure to more accuratelyrepresent the weaker pairwise matches). Distance-based genometrees based on much less stringent thresholds have recentlybeen published by Grishin et al. (17) and Wolf et al. (44),although PDSs were not identified or removed.

We identified 4,331 ORFs (10.1% of the 42,925 analyzable ORFs)that have patterns of BLASTP matches which are significantlydifferent from the patterns of their host genomes and wouldthus probably specify incongruent trees. Removing these ORFsfrom the analysis affected the topology of the genome tree verylittle, although it did improve the confidence in a number ofsubtrees as assessed by the nonparametric bootstrap method.We therefore concluded that both according to the proportionsof shared ORFs and according to sequence divergence, genomephylogeny largely agrees with the canonical view of prokaryoteevolution based on 16S rRNA gene sequences. It is especiallynoteworthy that restricting the analysis to ORFs that were foundnot to be discordant produced so few topological changes; thisdemonstrates that neither the set of nonanalyzable ORFs (47.1%of the ORFs) nor the 5.3% of the total ORFs identified as PDSsat P < 0.05 is specifically biased toward a preferred alternativetopology. Removal of PDSs improved bootstrap support most noticeablyin the ${gamma}$ - and ß-proteobacterial subtree, although thecomparison was imperfect, as there was limited or no room forimprovement in several regions of the tree.

Trees based on overall pairwise genomic distances do not, however,agree completely with the 16S rRNA tree. This is especiallytrue for the archaeal subtree, since in all our genome treesthe genome of the crenarchaeote A. pernix grouped with the genomesof euryarchaeotes. Euryarchaeota are paraphyletic in other genometrees as well (Fig. 3 and 5 of reference 44) and in a multigenetree based on sequences of ribosomal proteins (Fig. 6 of reference44), while Crenarchaeota are paraphyletic in trees based onradA (40). Thus, single-gene trees, such as 16S rRNA trees (5,35), are increasingly isolated in suggesting that Euryarchaeotaand Crenarchaeota are monophyletic. The basal position of thegenome of Halobacterium sp. (Fig. 2, 4, and 5) is unexpectedbased on small-subunit ribosomal DNA analysis (Fig. 3) but occursin the distance-based genome tree of Wolf et al. (Fig. 8 ofreference 44) and in a multigene tree based on concatenatedribosomal proteins (Fig. 6 of reference 44). We suspect thatthis could be an artifact of the high G+C content of this organismand the resultant systematic bias in the amino acid compositionof the proteins (15) or an artifact of the elevated contentof acidic amino acids that help stabilize proteins in the presenceof the high concentrations of intracellular salt characteristicof halobacteria (10). Haloarchaeal genomes may also containsignificant numbers of laterally transferred genes (6). It ismore difficult to explain why the genome of T. acidophilum branchesmore basally than expected, although this occurs in other genometrees (Fig. 3 and 5 of reference 44) and in the tree based onribosomal proteins (Fig. 6 of reference 44). Methanogens aremonophyletic in our trees (Fig. 2, 4, and 5) and other genometrees (Fig. 3 of reference 17; Fig. 3 and 5 of reference 44)but not in the trees based on small-subunit ribosomal DNA (5,40) (Fig. 3), radA (40), or ribosomal proteins (Fig. 6 of reference44).

Among the bacteria, T. maritima and A. aeolicus appear on thedeepest (most basal) branches in the trees based on 16S rRNA(5, 35) (Fig. 3) and ribosomal proteins (Fig. 6 of reference44) and in some (Fig. 3 and 5 of reference 44) but not all genometrees. In our analyses, the genome of T. maritima always appearedto be the most basal genome, but the genome of A. aeolicus wassecond deepest only in our distance-based trees (with or withoutremoval of discordant sequences). The numbers of and degreesof similarity between proteins in one or both of these hyperthermophilicbacteria and some archaea have opened a debate about commonancestry versus LGT (2, 3, 23, 28).

Spirochete and chlamydial genomes form a single clade in ourdistance-based genome trees (Fig. 4 and 5), in the distance-basedgenomes tree of Grishin et al. (Fig. 3 of reference 17) andWolf et al. (Fig. 5 of reference 44), and in a multigene ribosomalprotein-based tree (Fig. 6 of reference 44). They do not grouptogether in genome trees based on proportions of shared genes(Fig. 2 of reference 42; Fig. 3 of reference 44) (Fig. 2) orin many 16S rRNA-based trees (34) (Fig. 3). No genome treessupport monophyletic grouping of all Firmicutes (low-G+C-contentand high-G+C-content gram-positive organisms), as some 16S rRNAtrees do (Fig. 3). Previously, the low-G+C-content gram-positivebacteria (Bacillus-Clostridium group of Firmicutes) were polyphyleticin genome trees based on proportions of shared genes (Fig. 2of reference 42; Fig. 3 of reference 44) but monophyletic indistance-based genome trees (Fig. 3 of reference 17; Fig. 5of reference 44), 16S rRNA-based trees 34) (Fig. 3), and concatenatedribosomal protein gene-based trees (Fig. 6 of reference 44).These organisms were monophyletic in all of our genome trees(Fig. 2, 4, and 5).

Our genome trees are the first trees to resolve the Proteobacteriaas a stable monophyletic group, in agreement with 16S rRNA trees(34) (Fig. 3) and concatenated ribosomal protein trees (Fig.6 of reference 44); the levels of bootstrap support were high(99 and 100% before and after removal of discordant sequences,respectively). Our proportion-based genome tree (Fig. 2) maybe unique in resolving the ß-proteobacteria as a sisterlineage of the ${gamma}$ -proteobacteria; in our distance-based trees(Fig. 4 and 5), as in the trees based on16S rRNA (34) (Fig.3) and ribosomal proteins (Fig. 6 of reference 44), Neisseriagroups with the ${gamma}$ -proteobacteria.

Details aside, many genome trees are substantially similar tothe 16S rRNA gene tree and to many other single-gene trees andcontain physiologically coherent groups largely consistent withmodern prokaryote systematics. This finding is obviously consistentwith the Darwinian model of descent along genealogical lineages,as on a bifurcating tree, but in and of itself it does not provethat prokaryotes have evolved mostly by traditional verticaldescent. Doolittle (8, 9; personal communication) has suggesteda scenario of genome evolution in which LGT plays a predominantrole and lineages owe their existence to selective sharing ofgene pools that might be constrained by environmental, physiological,and other nongenealogical factors. In this scenario, distancessuch as those underlying the construction of genome trees mightreflect the frequency of LGT, not the time since a common ancestor.The genes that confuse genome phylogenies are a testament tothis lateral exchange; we explore them further elsewhere (39).Thus, it remains to be determined whether the topologies ingenome trees reflect Darwinian evolutionary lineages or areartifacts of an elaborate network of differential lateral geneticexchange. We strongly urge that our results be interpreted inthis context.

ACKNOWLEDGMENTS

We are grateful to A. St. Jean for his efforts in developingthe NeuroGadgets Inc. Bioinformatics Web Service and to membersof the Evolutionary Biology Program of the Canadian Institutefor Advanced Research for helpful comments.

This work was funded by the Natural Sciences and EngineeringResearch Council of Canada and by the Institute for MolecularBioscience, University of Queensland. We are grateful to G.Drouin for financial support of G.D.P.C.

FOOTNOTES

* Corresponding author. Mailing address: The Institute for Molecular Bioscience, The University of Queensland, Brisbane, Qld 4072, Australia. Phone: 61-7-3365-1160. Fax: 61-7-3365-4388. E-mail: m.ragan{at}imb.uq.edu.au.

REFERENCES

Top