Genome Phylogenies Indicate a Meaningful a-Proteobacterial Phylogeny and Support a Grouping ...

12 Pages·2005·0.31 MB·English

Genome Phylogenies Indicate a Meaningful a-Proteobacterial Phylogeny
and Support a Grouping of the Mitochondria with the Rickettsiales
David A. Fitzpatrick, Christopher J. Creevey, and James O. McInerney
Department of Biology,National University of Ireland,Maynooth, CountyKildare, Ireland
Placementofthemitochondrialbranchonthetreeoflifehasbeenproblematic.Sparsesampling,theuncertaintyofhow
lateralgenetransfermightoverwritephylogeneticsignals,andtheuncertaintyofphylogeneticinferencehaveallcontributed
to the issue. Here we address this issue using a supertree approach and completed genomic sequences. We first
determinethatasensiblea-proteobacterialphylogenetictreeexistsandthatitcanconfidentlybeinferredusingorthologous
genes.Weshowthatcongruenceacrosstheseorthologousgenetreesissignificantlybetterthanmightbeexpectedbyran-
domchance.Thereissomeevidenceofhorizontalgenetransferwithinthea-proteobacteria,butitappearstoberestrictedto
aminorityofgenes(;23%)mostofwhom(;74%)canbecategorizedasoperational.Thismeansthatplacementofthe
mitochondrionshouldnotbeexcessivelyhamperedbyinterspeciesgenetransfer.Wethenshowthatthereisaconsistently
strongsignalforplacementofthemitochondriononthistreeandthatthisplacementisrelativelyinsensitivetomethod-
ological approachordata set.Aconcatenatedalignmentwascreatedconsistingof 15mitochondrion-encoded proteins
thatareunlikelytohaveundergoneanylateralgenetransferinthetimelineunderconsideration.Thisalignmentinfers
thatthesistergroupofthemitochondria,forthetaxathathavebeensampled,istheorderRickettsiales.
Introduction
Proteobacteria or the purple bacteria are one of the Anyattempttodeterminewhichextanta-proteobacte-
largestknowndivisionswithinprokaryotes.Basedonanal- riumisthesistergroupofthemitochondriadependsonthe
ysis of the evolution of 16S rRNA sequences, the proteo- hypothesis that there isa robust and meaningful a-proteo-
bacteriadivisionwasfirstcircumscribedbyCarlWoeseand bacterial phylogeny. This in turn implies that within this
coworkers (Woese et al. 1985; Woese 1987). Subsequent grouphorizontalgenetransfer(HGT)hasbeensoinfrequent
16S rRNA analysishas led to the proteobacteria being di- that a meaningful species phylogeny can be derived. The
videdintofivesubdivisions(a,b,d,c,ande;Woese1987). alternative that HGT is rampant and no phylogeny exists
Theproteobacterialgroupisanimportantdivisionasit would mean that attempting to ascertain the sister group
includesknownanimal,human,andplantpathogens.Fur- tothemitochondriaispointless.Up until nowsingle-gene
thermore,thisdivisionhasplayedacrucialroleineukaryotic phylogenies (especially small-subunit rRNA-based ones)
cell origin and evolution through endosymbiosis. The hy- have established many of the accepted relationships be-
pothesispertainingtoanendosymbioticoriginofthemito- tween bacteria. Single-gene analyses are dependent on
chondria was first proposed in the 19th century (Altman the gene having an evolutionary history that reflects that
1890)andlaterpopularizedbyMargulis(1981).Today,there oftheentireorganism.Abetterapproachwouldbetocom-
is strong evidence to support the hypothesis that present-
bineinformationfrommanygenes,andtothisend,supertree
day mitochondria were once free-living a-proteobacteria
analysisisalogicalmethodtoemployasitaimstocombine
(Anderssonetal.1998;Lang,Gray,andBurger1999;Ogata
informationfrommanyindividualgenes.Anotherapproach
etal.2001).However,theexactpositionofthemitochondria
wouldbetoconcatenategenes.Concatenationofsequence
withinthea-proteobacteriaisstilldebated(Wuetal.2004).
data encounters problems when there is some evidence of
Anumberofanalyseshasplacedthemitochondriaamong
HGT.Theseproblemsarecomparabletothoseencountered
the order Rickettsiales (Gupta 1995; Lang, Gray, and
when reconstructing phylogenies with genes that have
Burger 1999), while others propose that mitochondria are
recombined(SchierupandHein2000).Conversely,because
more closely related to the Rickettsiaceae family and
ofthewayinwhichsupertreessummarizetaxonomiccon-
Rickettsia prowazekii in particular to the exclusion of the
gruence,theylimittheimpactofindividualgenetransferson
Ehrlichia and Anaplasma genera (Karlin and Brocchieri
theglobaltopology(Escobar-Paramoetal.2004).Itisthen
2000;Emelyanov2003).OncompletionoftheWolbachia
possible using supertree methods to conduct a posteriori
pipientiswMelgenomesequence,Wuetal.(2004)reported
analyses ofcongruence between input trees andsupertree.
strongsupportforagroupingofWolbachiaandRickettsia
The primary reason for the prerequisite of a robust
as a sister group to the exclusion of the mitochondria. A
supertree is to account for possible HGT events. HGT
recent analysis that utilized 55 individual and 31 concate-
amongbacteriahasbeenshowntobeamajorsourceofge-
nated genes encoded in Reclinomonas americana and
neticvariation,andreportedincidencesofHGTarecontin-
Marchantia polymorpha mitochondrial genomes by Esser
ually increasing (Martin et al. 1998; Wolf, Aravind, and
et al. (2004) concluded that based on the available data
Koonin 1999; de la Cruz and Davies 2000; Brown 2003;
‘‘Rhodospirillum rubrum comes as close to mitochondria
Kinsella et al. 2003). By investigating congruence across
asanya-proteobacteriuminvestigated’’(Esseretal.2004).
ortholog-derivedphylogenetictrees,wecangaugehowof-
tenHGThasoccurred.Ifweknewhowmucherrorwemight
Key words: a-proteobacterial phylogeny, supertree, mitochondria, expect, then we would know how much confidence we
removingfast-evolvingsites.
mightinvestinourconclusions.Recently,therehavebeen
E-mail:[email protected].
suggestionsthatHGT,ingeneral,hasnotbeensoseverethat
Mol.Biol.Evol.23(1):74–85.2006 ithaswipedoutphylogeneticsignalamongcloselyrelated
doi:10.1093/molbev/msj009
AdvanceAccesspublicationSeptember8,2005 species (Daubin, Gouy, and Perriere 2001; Daubin, Lerat,
(cid:1)TheAuthor2005.PublishedbyOxfordUniversityPressonbehalfof
theSocietyforMolecularBiologyandEvolution.Allrightsreserved.
Forpermissions,pleasee-mail:[email protected]
GenomePhylogeniesIndicateaMeaningfula-ProteobacterialPhylogeny 75
and Perriere 2003), but ortholog trees contain enough GENERATOR chooses the best amino acid substitution
conflict that early prokaryotic evolution cannot be repre- modelfromatotalof80possiblemodels.Bootstrapresam-
sentedeffectivelywithauniquephylogenetictree(Creevey pling was carried out 100 times on each alignment, and
et al.2004). usingthemostappropriateoftheavailablemodels,phylo-
In this study, we examine the nature and extent of geneticinferenceswereconstructedusingthesoftwarepro-
HGT among the a-proteobacteria, we identify widely dis- gramPHYML(GuindonandGascuel2003).Theresultsof
tributed genes that appear to have a history that does not these analyses were summarized using the majority-rule
includeHGT,weusethesegenesinordertoinferthesister consensus method.
grouprelationshipbetweenthemitochondriaandmembers
ofthea-proteobacteria,andwediscusstheimplicationsof
Supertree Reconstruction
these inferences.
Inordertorepresentthevariabilitywithinthedataset,
thebesttreesfromeachofthe100bootstrapreplicatesfrom
Materials and Methods all 418 single-gene families were combined as the source
Sequence Data dataforasupertreeanalysis(atotalof41,800inputtrees).
Thebestsupertreewasfoundfollowingaheuristicsearchof
The bacterial database used in this analysis consisted
treespaceusingthemostsimilarsupertreeanalysis(MSSA)
of 13 completely sequenced a-proteobacteria genomes,
(Creevey et al. 2004) as implemented in CLANN 2.0.2
i.e., Sinorhizobium meliloti, Agrobacterium tumefaciens,
(CreeveyandMcInerney2005).UsingCLANN,100boot-
Wolbachia,R.prowazekii,Brucellamelitensis,Caulobacter
strap resamplings were also carried out on the 41,800
crescentus, Mesorhizobium loti, Anaplasma marginale,
source tree data set to assess the level of support for the
Ehrlichia ruminantium, Bradyrhizobium japonicum,
internal branches. We tested for the presence of signal in
Rhodopseudomonas palustris, Bartonella quintana, and
the data set by performing the Yet Another Permutation-
Silicibacter pomeroyi and the partial sequence data of
Tail-Probabilitytest(YAPTP)randomizationtest(Creevey
Novosphingobium aromaticivorans, Rhodobacter sphaer-
andMcInerney2005), whichteststhenullhypothesisthat
oides, and R. rubrum. Magnetospirillum magnetotacticum
congruencebetweentheinputtreesisnobetterthanrandom
MS-1 homologs were used in three incidences in place of
(Creevey and McInerney 2005). In order to examine the
R. rubrum (explained later). Sequence data for the partial
behaviorofperfectdata,wegeneratedinputtreesthatwere
genomes were supplied by Department of Energy Joint
completelycongruentwiththesupertree(idealizeddataset)
GenomeInstitute(http://www.jgi.doe.gov/).
usingthe‘‘generatetrees’’commandinCLANN.Foradetailed
Inadditiontothea-proteobacterialgenomes,weused
explanationoftheabovemethods,seeCreeveyetal.2004.
the complete mitochondrial genomes of R. americana,
Malawimonas jakobiformi, and M. polymorpha as well as the
genome sequences of Neisseria meningitidis MC58 and Shimodaira–Hasegawa Tests
Escherichia coli 0157:H7 which were used as outgroups.
Toassesswhetherdifferencesintopologybetweenthe
supertree and individual gene trees are no greater than
a-Proteobacterial Phylogeny
expectedbychance,weperformedShimodaira–Hasegawa
For supertree construction, homologous sequences (SH) tests (Shimodaira and Hasegawa 1999) using Tree-
were identified by performing an all-against-all search of Puzzle5.1(Schmidtetal.2002).Genetreesvariedinsize
thebacterialdatabaseusingtheBlastPalgorithm(Altschul from4to16taxa;therefore,whentherewerefewerthan16
et al. 1997) with a cutoff expectation (E) value of 10(cid:1)7. sequencesinanyinputdataset,thesupertreewasappropri-
Only those homologous families where every member atelyprunedsothatitcontainedthesameleafsetthatwas
foundeveryothermember(andnothingelse)andweredis- found in the individual gene trees. The underlying amino
tributed across at least four taxa, and when present, were acid alignment from which the input tree was derived
alwaysinsinglecopy,wereretained.Thisconservativeap- was used for each SH test.
proachhasbeendesignedtominimizetheinadvertentanal-
ysisofparalogsorsplicedgenes.Intotal,539familiesmet
Mitochondrial Relationships
these criteria. The individual protein families were then
aligned using ClustalW 1.81 (Thompson, Higgins, and UsingBlastP,withacutoffEvalueof10(cid:1)7,thehomo-
Gibson 1994) using the default settings. All alignments logs that were common to R. americana, M. jakobiformi,
were corrected for obvious alignment ambiguity using andM.polymorphamitochondriawereusedtofindhomo-
the alignment editor Se-Al 2.0a11, and permuation-tail logsamongthebacteria(Altschuletal.1997).Insomein-
probability (PTP)tests were carried outto test for an evo- stances, more than one match per genome was found. In
lutionary signal that was significantly better than random these cases, the best match to the mitochondrial query
noise (P , 0.01). We found that 121 alignments failed wasretained.Thisapproachallowseacha-proteobacterium
the PTP test and the remaining 418 alignments were used to be as similar to mitochondria as possible at the level of
for phylogenetic analysis. sequence similarity and follows the procedure of Esser
Appropriate protein models were selected for each et al. (2004). Appropriate homologs from cox1, cox2, and
of the 418 single-gene families within the supertree data cox3 genes could not be located for R. rubrum; therefore,
set using the software program MODELGENERATOR following the procedure taken by Esser et al. (2004), M.
(http://bioinf.nuim.ie/software/modelgenerator). MODEL- magnetotacticum homologs were used instead, given that
76 Fitzpatricketal.
Rhodospirillum and Magnetospirillum are almost always Table1
well-supported sister taxa (Esser et al. 2004). In total, 28 Mitochondrion-Encoded ProteinsWhosePhylogenyIs
genefamilieswerefoundtocontainall21taxa(N.menin- not SignificantlyDifferent (P>0.05)tothe Proposed
gitidis and E. coli outgroups, 3 mitochondria, and 16 a- a-Proteobacterial Supertree According toSH Test
proteobacteria).Individualmitochondrion-encodedprotein Gene LogDet BP ML BP Bayesian Prob Sites
families were aligned using ClustalW 1.81 (Thompson,
cox3 Rick1og 72 Rick1Bar 80 Rick1Bar 0.95 396
Higgins,andGibson1994)usingthedefaultsettings.Gaps rpl5 Rick 84 Rick1og 52 Rick 1 200
createdintheaminoacidalignmentswereinsertedintothe rps14 Rick 70 Ricka 70 prow 0.93 102
nucleotide sequences to produce codon-based nucleotide rps11 Rick 87 Rick 71 Rick 0.97 132
nad3 Rick 68 Ricka 97 Ricka 0.97 130
alignments using the program putgaps (http://bioinf.nuim.
rps2 Rick 62 Rick 60 Rick 1 336
ie/software/putgaps). All alignments were corrected for
rps3 Rick 88 prow 80 prow 0.99 253
obvious alignment ambiguity using the alignment editor rps8 Rick1og 97 Rick1og 51 prow 1 133
Se-Al2.0a11. rps12 Rick 98 Rick 96 Rick 1 123
Phylogenetic relationships for the 28 mitochondrion- rpl6 Rick 82 Rick 74 Rick 0.95 189
rps1 Rick 79 Rick ch Rick 1 591
encodedgenes(mitochondrialdataset)werereconstructed
rps13 Rick1og 70 Rick 82 Rick1og 0.97 121
using a variety of methods. Inference of relationships was rps19 Rick 73 Rick 96 Rick 0.99 94
carriedoutusingtheBayesianapproachimplementedinthe nad4l (Rick1og) all Rick1Bar 72 Rick 0.99 102
MrBayes v3.0B4 software (Huelsenbeck and Ronquist rpl16 Rick1og ch Rick 97 Rick 0.97 137
2001). Codon-based nucleotide alignments were analyzed
NOTE.—Rick,Rickettsiales;prow,Rickettsiaprowazeki;Bar,Bartonellaquin-
withtheassumptionthatamong-siteratevariationcouldbe tana;og,Escherichiacoli1Neisseriameningitidis;BP,bootstrapsupportvalues;
describedusingadiscreteapproximationtothegammadis- Prob,Bayesiancladeprobabilities.TheoptimumtopologyaccordingtotheSHtest
tribution (four categories of sites), and a proportion of the areinboldface.ThesistertaxaofthemitochondrionaccordingtoLogDet,ML,and
Bayesianreconstructionmethodsareshownincolumns2,4,and6,respectively.
alignment was constrained to be invariant. The shape pa-
CorrespondingbootstrapsupportvaluesandBayesiancladeprobabilitiesareshown
rameter of the gamma distribution and the hypothesized incolumns3,5,and7.
proportion of invariant sites were allowed to vary through a ExceptAnaplasmamarginale.
the Markov Chain Monte Carlo (MCMC) chain. In total,
four MCMC chains were run for 1.5 million generations;
phylogenetic trees were sampled every 100th generation. phylogeny is not significantly different (according to the
Plots of the likelihood values revealed that the burn-in SH test) to the proposed a-proteobacteria supertree. The
phase took at most 200,000 generations. Therefore, reasoningbehindthisstrategyisthatthesegenesprobably
200,000 trees were discarded for all alignments. Clade have not undergone HGT. This concatenated alignment
probabilities for each phylogeny were determined using consisted of 15 genes (table 1) and in total had 3,039
the sumt command of MrBayes v3.0B4. Maximum likeli- alignedaminoacidpositions;removalofgappedpositions
hood(ML)phylogeniesbasedonaminoacidsequencesde- reduced the alignment to 1,781amino acids. Phylogenetic
rived from the codon-based nucleotide alignments were relationshipsforthetaxawithinthisalignmentwerederived
reconstructed using PHYML, with the model of sequence fromLogDetdistancesandtheMLcriterion.MLtreeswere
substitution and estimated parameters governing rate vari- reconstructed using PHYML using the appropriate model
ation across sites determined by MODELGENERATOR. ofsequencesubstitutionaccordingtothehierarchiallikeli-
Additionally,neighbor-joining(NJ)treesbasedonnucleo- hoodratiotestimplementedinMODELGENERATOR.An
tide LogDet distances (Lockhart et al. 1994) were recon- NJ tree based on amino acid LogDet distances (Lockhart
structed using PAUP* 4.0b10 (Swofford 1998). Constant et al. 1994) was reconstructed using LDDIST (Thollesson
and third codon positions were removed from all align- 2004), the fraction of invariant sites was estimated by the
ments for the LogDet analyses. Support for relationships method of Sidow, Nguyen, and Speed (1992), and these
weredeterminedbycarryingout100bootstrapresamplings were excluded. Support for groups on trees were deter-
ofthedata,constructingphylogenetictrees,andsummariz- minedusingthebootstrapresamplingtechnique.Splitsim-
ing the results using a majority-rule consensus method. plied by sites in the concatenated alignment were found
Phylogenetic trees derived from the equivalent protein usingtheLogDetdistancesdescribedaboveandNeighbor-
alignmentswerealsoreconstructedusingMLandBayesian Net(BryantandMoulton2004).Thesewererepresentedas
methods, and in all cases nucleotide- and protein-derived aplanargraphusingtheSplitsTreesoftware(Huson1998).
treeswerecomparable.Wecouldnotreconstructphyloge- Phylogenetic networks permit the representation of con-
netic trees using the LogDet transformation on protein flicting signal or alternative phylogenetic histories (Fitch
sequences for the majority of our 28 mitochondrion- 1997).Whentheevolutionaryhistory ofgenes isnottree-
encoded genes because there were instances where the like, it is necessary to use phylogenetic networks (Bryant
determinants were zero and the distances undefined (see and Moulton 2004). Even when the underlying history is
Thollesson 2004). treelike, parallel evolution and sampling error may make
it difficult to determine a unique tree. In these cases, net-
works provide a valuable tool for representing ambiguity
or for visualizing a space of feasible trees (Bryant and
Concatenated Mitochondrial Data Set
Moulton 2004).
Aconcatenatedalignmentwasconstructedconsisting Wealsosubjectedthealignmenttoaspectralanalysis.
onlyofthosemitochondrion-encodedgeneswhoseinferred Spectralanalysisisaquantitativemethodthatissuitedfor
GenomePhylogeniesIndicateaMeaningfula-ProteobacterialPhylogeny 77
(A) R. americana
M. polymorpha
Wolbachia
R. prowazekii
E. coli
R. rubrum N. meningitidis
N. aromaticivorans
C. crescentus
M. loti S. meliloti R. sphaeroides
B. melitensis
A. tumafaciens
(B) R. americana
R. prowazekii
M. polymorpha
Wolbachia
N. meningitidis
R. rubrum
N. aromaticivorans
A. tumafaciens
R. sphaeroides S. meliloti
C. crescentus M. lotiB. melitensis E. coli
FIG.1.—ComparisonoftwophylogeneticNeighborNets.Theunderlyingalignmentconsistsof31concatenatedmitochondrion-encodedgenesandis
identicalinsizeandcontenttothealignmentanalyzedbyEsseretal.(2004).Thealignmenthashadcategoriesoffast-evolvingsitesremoveduntilthe
alignmentpassesav2test(P,0.05)foraminoacidhomogeneity.Thetopphylogeneticnet(A)isderivedfromanaminoacidalignmentwithfast-
evolvingsitescategorizedusingthemethodofHansmannandMartin(2000).Thesecondphylogeneticnet(B)hashadfast-evolvingsitescategorizedby
agammadistributionremoved.FollowingthemethodologyofEsseretal.(2004),distancesweredeterminedusingtheLogDettransformationand
invariantsiteshavebeenremoved.Network(A)isidenticaltothatpresentedbyEsseretal.(2004).Innetwork(B)theRickettsialesareinferredto
bethesistertaxonofthemitochondrion.Differencesinnetworktopologiesaresolelytheresultofdifferentmethodsusedtostripfast-evolvingsites.
testing alternative phylogenetic hypotheses when relation- tionally homogenous alignment of 574 amino acids was
ships among taxa are controversial (McCormack et al. found. The second method is implemented in Tree-Puzzle
2000;Winchelletal.2002).Aspectralanalysisonprotein 5.1 and assumes that there are eight categories of sites.
LogDet distances creates a complete array of bipartitions Ratevariationacrossthesesitesisassumedtofollowadis-
andsplitsinthedatawhichincludeseverypossiblegroup- crete gamma distribution. Again, fast-evolving categories
ingamongthetaxaused.Thestrengthofthismethodisthat of sites were iteratively removed until an alignment of
it works independently of any one particular tree (Lento 651aminoacidpositionsthatpassedthev2testforhomo-
et al. 1995). geneityofaminoacidcompositionwasfoundforthe15-gene
AccordingtoTree-Puzzle5.1,theconcatenatedalign- data set. Phylogenetic hypotheses for the homogeneous
ment fails the v2 test for homogenous amino acid compo- alignments were generated using the ML framework de-
sition(P,0.05).Thehighlyvariablesiteswereiteratively scribed earlier.
removedfromtheconcatenatedalignmentusingtwodiffer- To account for problems associated with deep level
entmethodsinordertogivethelargestpossiblealignment relationships, the amino acid alignment was recoded into
inwhichallsequencespassedthev2testforhomogeneityof thesixDayhoffgroups:C,STPAG,NDEQ,HRK,MILV,
aminoacidcomposition.Thereasonforutilizingtwometh- andFWY.Thisisbasedonthenotionthataminoacidsub-
odswasbecausewhenweanalyzedthedataofEsseretal. stitutionswithinthesixgroupswillbecommonandnoisy,
(2004)wefoundanalternativephylogeneticnetworktothe whilechangesbetweengroupswillberarerandsohaveless
one they proposed (fig. 1). The reason for this alternative saturation (Hrdy et al. 2004). The recoded alignment was
inference was due to the fact that we had used a different analyzed using the Bayesian criterion implemented in
method to remove fast-evolving sites. MrBayes v3.0B4 (Huelsenbeck and Ronquist 2001). The
The first method described by Hansmann and Martin modelusedhasa636generaltime-reversibleratematrix.
(2000)scoressitesdependingonthenumberofaminoacid Among-site variation in evolutionary rate was modeled
character states that are found at that site. This effectively withanallowanceforaproportionofsitestobeinvariable
assumes that there is a star phylogeny with equal branch andfourcategoriesofvariablesiteswiththeirratesmodeled
lengthsandisthemethodusedbyEsseretal.(2004).Sites byadiscreteapproximationtothegammadistribution.The
withhighscoreswereiterativelyremoveduntilacomposi- analysis used an MCMC chain that ran for 2 million
78 Fitzpatricketal.
generationssampledevery100thgeneration;thefirst5,000 compatiblewiththesupertree(idealdata),andalsotheinput
sampleswerediscardedastheburn-in.Supportforgroups treeswerecompletelyrandomized100differenttimesinor-
on the tree were determined using the sumt command of dertocarryouttheYAPTPtest.Thescoresof10,000ran-
MrBayes, which produces a majority-rule consensus tree dom supertrees were then obtained based on the real
generated from bipartitions found most frequently during (unperturbed), ideal, and randomized input trees, and the
the MCMC runs. distributionsofthesesupertreescoreswereassessed.Inad-
dition,wesoughttodeterminehowmanyperturbationsper
treewereneededtogivetheidealdatathekindsofcharacter-
Results isticsthatweobserveintherealdataandhowmanypertur-
a-Proteobacteria Supertree bationsmightberequiredtocompletelyrandomizetheinput
trees.Inthisway,wedeterminedhowclosetherealdatawas
Totestwhetherthereisameaningfula-proteobacterial
to ideal data andhow close it was to random data.
phylogeny,adatasetconsistingof16a-proteobacterialge-
The range of supertree scores for the real data varied
nomesor60,640individualcodingsequenceswasutilized.
from33,196(best)to92,027(worst)(fig.2).Ingeneral,the
Weidentified418single-genefamiliesthatmetourcriteria
scoresoftheworstsupertreesfromallthreedatasetswere
ofbeingpresentinsinglecopyinatleastfourgenomesand
similar.Thedistributionofthescoresoftheoptimalsuper-
that have phylogenetic signal according to the PTP test.
trees from the YAPTP test is centered on 70,400 (6300).
Aminoacidsequenceswerealignedresultinginacombined
This indicates that congruence across the input trees is
length of 140,314 aligned positions. Bootstrapresampling
greater than expected by random chance (Creevey et al.
followedbyphylogeneticinferenceresultedin100MLtrees
2004). The ideal data generated in this study represents
foreachfamily.Usingthese41,800treesasinputdatafor
a scenario where there is complete compatibility between
supertreeanalysis,heuristicsearchesofsupertreespacewere
the input trees and supertree in figure 2. The distribution
carriedoutandtheoptimalsupertreewasidentified(fig.2).
of tree scores for this ideal data set was heavily skewed
One hundred bootstrap resamplings of the input trees was
to the left, with a skewness value of (cid:1)0.8. The real data
carriedout,andthesupportvaluesoftheinternalbranches
wereslightlymoreskewedtotheleftwithaskewnessvalue
wereplacedonthesupertreeinfigure2.
of (cid:1)0.11. The randomized data had a skewness value of
(cid:1)0.02,whichisclosetozero,asexpected.Allthreedistri-
How Compatible Are the Input Trees?
butions had a standard deviation (SD) of 0.02.
Anidealizeddatasetwascreatedwherethetopologyof Wethentooktheidealdatasetandapplyingthesub-
theinputtreeswaschangedtomakethesetreescompletely tree pruning regrafting (SPR) method on the input trees
FIG.2.—Theshapeofsupertreespace.Indiamondsarethescoresfor10,000randomsupertreeswhenscoredagainstanidealsourcetreedataset.The
besttreefromtheidealdistributionhasascoreof0andtheskewnessoftheidealdistributionis(cid:1)0.08.Thetrianglesarethescoresof10,000random
supertreeswhencomparedtotherealsourcetreedataset.Theskewnessofthisdistributionis(cid:1)0.11.Heuristicsearchesofsupertreespaceusingthereal
sourcedatarevealedthatthebestsupertreehadascoreof33,196,andwhenbootstrapped,nearlyeveryinternalbranchofthetreehad100%bootstrap
supportvalue.Thesquaresarethescoresof10,000randomsupertreeswhencomparedtoarandomizedsetofsourcetrees,thisdistributionhadaskewness
of(cid:1)0.02.TheSDoftheskewnessfortheideal,real,andrandomizeddatawas0.02.Thecirclesdemonstratetheresultsof100YAPTPtests.
GenomePhylogeniesIndicateaMeaningfula-ProteobacterialPhylogeny 79
(equivalenttointroducinganHGTevent),weincrementally these 15 mitochondrion-encoded genes were constructed
perturbedtheideal data set. TheseSPR events wereintro- for the codon-based nucleotide alignments using LogDet,
duced,andateachstagethebestsupertreewasfoundandits ML, and Bayesian methods. All three methods revealed
scoreforthedatasetassesseduntilthescore,accordingto a general consensus regarding the possible sister group
theMSSAmethod, was thesame as either thereal orran- of mitochondria (table 1). The majority of these 15 gene
domizeddata.Thisgaveanindicationofthenumberofper- trees placed the members of the Rickettsiales as the sister
turbations required to make ideal data as noisy as the real groupofthemitochondria.WecarriedoutavarietyofSH
(orrandom)dataset.Eachofthesesimulationswascarried tests using the Whelan and Goldman (2000) model of
out100times.Fromthe100simulations,themeannumber amino acid substitution. For each gene, the underlying
ofSPReventspertreerequiredtotransformidealdatainto alignment and thethree proposed topologies derived from
the equivalent of the real data set was 1.6 (with an SD of the above three methods were used for each gene family.
0.07).Totransform theidealdataintodatawithanequiv- An additional ‘‘constrained’’ tree where R. rubrum was
alent level of incongruence as random data required 5.0 placedasasistergrouptothemitochondria(recreatedusing
HGT events per tree (with an SD of 0.28). MrBayes)wasalsoincludedintheanalysis.Thereasoning
Taken together, it can be seen that the real data are behindthisapproachisthatR.rubrumhasbeensuggested
significantly different than random data (according to the as a possible sister group to the mitochondria (Esser et al.
YAPTPtest)andbytwomeasuresitismoresimilartoideal 2004). According to the SH test, the constrained tree was
data than it is to random data. significantly (P . 0.05) worse than the three alternative
Ofthe418inputtrees,SHtestsindicatedthat98(23%) competing topologies in all cases (results not shown) and
whosetopologycouldnotbeinducedbytheappropriately wasthereforenotconsideredforfurtheranalysis.Examin-
prunedsupertree,describedtheirunderlyingalignmentssig- ingtheoptimumtopologiesaccordingtotheSHtestreveals
nificantlybetter(P,0.05)thandidthesupertree.Thedif- that in total 11 trees placed the Rickettsiales as the sister
ferences in topology could not be explained by chance grouptothemitochondria.OnetreeplacedtheRickettsiales
alone.Fortheremaining320trees,therewasnosignificant (except for A. marginale) as the sister group to the mito-
difference between the proposed supertree and underlying chondrion,anotherinferredthattheRickettsialescombined
tree (24% ofthe these are identical to theproposed super- withB.quintanawerethesistergroup,anothertreeplaced
tree).HGTeventsandhiddenparalogiesareamongthepos- the Rickettsiales and the outgroup together, and then an-
sible reasons for the 98 families having an evolutionary othersplitseparatedthisgroupandthemitochondriafrom
history that issignificantly different to that implied by the therestofthetaxa. The finaltree placedthemitochondria
supertree.Inanefforttoquantifythetypesofgenesthatdis- with R. prowazekii as sister taxa (table 1).
play significantly different topologies to the supertree, we
categorized genes as being either operational or informa-
Concatenated Alignment
tionalinfunctionbasedonthecriteriausedbyJain,Rivera,
and Lake (1999). According to the complexity hypothesis Concatenation of the 15 mitochondrion-encoded
(Jain,Rivera,andLake1999),wewouldexpecttoseeop- genes of interest gave an alignment of 3,039 amino acids
erationalgenesbeingexchangedatahigherratethaninfor- in length or 1,781 amino acids when positions containing
mationalgenes.Informational genes are members of large gapswereremoved.Concatenationofsequencedatahasthe
complex systems thereby making horizontal transfer of benefit of reducing stochastic effects related to short se-
these gene products less probable (Jain, Rivera, and Lake quencelength.Accordingtotheaminoacidcompositional
1999).Wefound87informationalgenesand96operational heterogeneitytest,therewassevereaminoacidbiaswithin
genes,theremaininggeneshaveyettobeassignedafunction the data set; therefore, the assumption of stationary amino
(i.e.,conservedhypothetical)orbelongtoacategorythatis acidfrequencies(GuandLi1996)wasviolated.Toaccount
neitheroperationalnorinformational.Intotal,18ofthe87 forthesebiases,theLogDettransformation(Lockhartetal.
informationalgenes(;20%)werefoundtohaveatopology 1994)wasusedtoinfertherelationshipsamongthea-pro-
thatissignificantlydifferenttotheproposedsupertree,while teobacteriaandthemitochondriaforthedatasetcontaining
50 of the 96 operational gene (;51%) trees differ signifi- no gaps. NJ trees constructed using LogDet distances de-
cantlytothesupertreeaccordingtotheSHtest.Assuming rivedfromtheaminoacidalignmentplacetheRickettsiales
that systematic or stochastic errors in phylogenetic recon- asthesistergrouptothemitochondriawith99%bootstrap
structionoftheinputtreesisnotfunctiondependentandalso support (fig. 3A). Phylogenetic hypotheses were also in-
thatthesupertreeistrulyrepresentativeoftherelationships ferred using the program PHYML using the appropriate
amongthea-proteobacteria,thiswouldindicatethattheop- modelofsequencesubstitutionselectedbyMODELGEN-
erationalgeneswithinthisdatasetexperiencesignificantly ERATOR. The ML topology was identical to the LogDet
(P , 0.001) more horizontal transfers or other anomalous topology and branch lengths and bootstrap support values
events than informational genes. were also comparable (fig. 5A). The NeighborNet of Log-
Det protein distances for this alignment also places the
Rickettsiales as the closest a-proteobacterial group to the
Mitochondrial Origin
mitochondria (fig. 3B). Comparing the supertree in figure
Individual Mitochondrion-Encoded Genes
2 to the tree built from the concatenated sequence data
The overlap between the 28 mitochondrion-encoded (fig. 3A), we see that the main orders such as the Rickett-
genesandthe301genesthatarecompatiblewiththesuper- siales, the Rhizobiales, and the Rhodobacterales are
tree was 15 genes. Phylogenetic hypotheses based upon strongly supported clades in both cases. In our supertree,
80 Fitzpatricketal.
(A) (B)
E. coli
b N. meningitidis N. aromaticivorans
S. pomeroyi
R. rubrum R. sphaeroides C. crescentus E. coli
N. aromaticivorans B. japonicum
N. meningitidis
98 M. loti RS.. pmaelulisltortiis
100 S. meliloti M. loti
B. melitensis
i
A. tumafaciens
100 100 B. melitensis B. quintana A. marginale
j 91 B. quintana R. rubrum E. ruminantium
e R. palustris
Wolbachia
100 99 100 B. japonicum
C. crescentus R. prowazekii
99 100 R. sphaeroides
f S. pomeroyi
R. americana
100 0.1
a M. jakobiformi
91 M. polymorpha R. americana
h 99
R. prowazekii
0.1 g M. jakobiformi
Wolbachia M. polymorpha
100 c
100 d E. ruminantium
100 A. marginale
(C)
a
b
c
d
e
f
g h i j
k
FIG.3.—TheMLphylogenyfortheconcatenatedalignmentisshownin(A).Bootstrapsupportsareshownforselectednodes.Theinferredtopology
isverysimilartoaninferencebasedonLogDetdistances(seefig.5A).TheRickettsialesgroupbesidethemitochondrionwith99%bootstrapsupport.A
NeighborNetoftheconcatenatedalignmentisalsodisplayedin(B);thisphylogeneticnetisbasedonproteinLogDetdistances.TheRickettsialesappearto
shareasplitwiththemitochondrion.Alentoplot(C)thatutilizestheaforementionedLogDetdistancesfortheconcatenatedalignmentisalsodisplayed.
Thetop40supportedsplits(excludingindividualtaxa)aredisplayed.Barsabovethexaxisrepresentfrequencyofsupportforeachsplit.Barsbelowthex
axisrepresentthesumofallconflictsagainstthecorrespondingsplitabovethexaxis.Lettersabovecolumnsrepresentparticularsplitsinthedata,and
thesehavebeenmappedontothephylogenetictreefordisplaypurposes.AttentionisdrawntosplithwhichincludesthemitochondrionandtheRick-
ettsialesandsplitkwhichincludesRhodospirillumrubrumandthemitochondrion.Theconflictforsplitkfaroutweighsanysupport.
C.crescentusandR.rubrumaregroupedassistertaxawith enatedalignmentplacestheseorphantaxaonthetreewith
poor support (50% bootstrap support). Novosphingobium very high levels of support. For example, C. crescentus is
aromaticivorans is grouped beside the Rhodobacterales, grouped beside the Rhodobacterales with 99% bootstrap
alsowith50%bootstrapsupport.Thesefindingsarenoten- support.
tirely surprising as these bacteria are lone members of the AspectralanalysisofLogDetdistanceswasperformed
orders Caulobacterales, Rhodospirillales, and Sphingomo- using Spectrum (http://taxonomy.zoology.gla.ac.uk/;mac/
nadales,respectively,inourtree.Contrastingly,theconcat- spectrum) to further investigate the phylogenetic signal
GenomePhylogeniesIndicateaMeaningfula-ProteobacterialPhylogeny 81
E. coli ing is real. The first test simply involved the systematic
N. meningitidis analyses of several data sets that only had a single long-
branch taxon included. This test was to see if the long-
R. rubrum
branchtaxawerestillattachingtothetreeinthesameplace,
N. aromaticivorans evenwhenotherlong-branchtaxawereabsent.Inallthree
1.00 A. tumafaciens cases, the long-branched taxa attached to the R. rubrum
1.00
branch with high bootstrap support (fig. 5).
S. meliloti
1.00 Thesecondtestoflong-branchattractioninvolvedthe
1.00 M. loti removalofthelong-branchtaxafromthetree(seefig.6)and
0.99 B. quintana their pseudorandom reinsertion 100 times. The taxa were
1.00
1.00 B. melitensis inserted into the tree as long as the point of insertion was
not beside another long-branch taxon. This resulted in the
B. japonicum
1.00 construction of phylogenetic trees with interspersed long
1.00 1.00 R. palustris and short branches, but where no long-branch taxa were
C. crescentus neighbors.Oncethetaxon-insertionprocesswascompleted,
1.00 S. pomeroyi theconcatenatedalignmentwasusedtodetermineappropri-
1.00 atebranchlengthsforalltaxa.Naturally,thebrancheswere
R. sphaeroides
quite long for the inserted taxa. We then simulated the
R. americana evolutionofaset ofsequencesforthisartificialtreeusing
1.00
M. jakobiformi Seq-Gen(RambautandGrassly1997),andusingthesame
1.00
methodsofanalysesthatwehaveusedpreviouslyonthecon-
M. polymorpha
0.99 catenateddataset,wereanalyzedthesimulatedsequencesto
R. prowazekii seehowoftenthelong-branchtaxawereincorrectlyplaced
1.00
Wolbachia beside each other (i.e., what was the strength of the long-
1.00 A. marginale branchattractioninadatasetofthisnature).Wefoundthat
1.00
0.1 eventhoughtherecouldbeslightdiscrepanciesbetweenthe
E. ruminantium
treeusedtosimulatethedataandthetreederivedfromthe
FIG.4.—InferredphylogenyshowsaBayesianconsensustreeofthe simulateddata,therewasneverasinglecasewherethelong-
15-geneconcatenatedalignment.Theanalysisusesageneraltime-revers-
branchtaxawereclusteredtogether.Althoughthistestcan-
iblesubstitutionmatrixestimatedfromaminoacidsequencesrecodedinto
notbetakenasdefinitiveproofthatlong-branchattraction
thesixDayhoffgroups.Among-siterateheterogeneitywasmodeledusing
afour-categorygammacorrectionwithafractionofinvariantsites.The aloneisnotresponsiblefortheirclusteringintherealdata
analysis used the MCMC strategy from MrBayes; parameters such as set,weatleastshowthatfordatasetswithsimilarcharacter-
thecompositionandsubstitutionratematrixwerefree.Posteriorprobabil- istics(composition,degreeofoverallevolutionarychange,
itiesforselectedbranchesareshownatnodes.
sequence length, mixture of long, and short branches)this
effectisnotsevereenoughtoresultincompletelyartifactual
present in our concatenated alignment. Lento et al. (1995) topologies.
have shown that a spectral analysis of LogDet distances We also recoded the amino acid data into the six
seemstoreducetheamountofrandomnoisecomparedwith groups of chemically related amino acids and analyzed the
anormalcharacter-basedspectralanalysis.AsSpectrumcan data using theBayesian criteria.Thisapproach hastheef-
onlyaccommodate20taxa,weremovedA.tumefaciensfor fect of shortening long branches and homogenizing the
this part of the analysis. With 20 taxa, thetotal number of amino acid composition among sequences (Hrdy et al.
possiblesplitsis220(cid:1)1or542,288.Ofthese,101splitsare 2004).Fromthisanalysis,wefoundthatthemitochondrion
supported by some evidence and the 40 most highly sup- and Rickettsiales are again strongly supported (fig. 4).
ported are represented by the bars shown in the top half Removing the fast-evolving sites from the 15-gene
of figure 3C. Attention is drawn to the split marked h, concatenated alignment using the method of Hansmann
the taxa within this split include the Rickettsiales and the and Martin (2000) until the reduced alignment passes the
mitochondria. The split marked k represents the grouping amino acid compositional heterogeneity test resulted in
of the mitochondria and R. rubrum as sister taxa. This re- analignmentof574aminoacidsinlength.TheNeighborNet
lationship is not observed in either our phylogenetic tree of LogDet protein distances for this shortened alignment
(fig. 3A) or phylogenetic network (fig. 3B). Furthermore, places the Rickettsiales as the closest a-proteobacterium
the degree of conflict for this split far outweighs the level group to the mitochondria. Bootstrap resampling was
ofsupport.Forcompleteness,wealsoperformedtheabove employed to assess levels of support for groups. This
analyses on the alignment with gap positions still present. showedthattheRickettsialesandthemitochondriaformed
The results of these analyses were in agreement with our acladewithrelativelylowbootstrapsupport(65%,results
observations in the gapless alignment. notshown).Usingthegammadistributiontomodelsite-rate
Lookingatourphylogenetictreeandnetwork(fig.3A variation in evolutionary rate and removing fast-evolving
and3B),itisobviousthatthetaxawithlonginternalbranches sites until the data passed the v2 test of compositional ho-
cluster together—the mitochondria, the Rickettsiales, and mogeneity resulted in an alignment of 651 amino acids in
the outgroups. We carried out two tests to investigate length. The NeighborNet of LogDet protein distances for
whetherthisgroupingisasystematicerror(long-branchat- this shortened alignment places the Rickettsiales beside
traction,alsoknownastheFelsensteinzone)orthegroup- themitochondriaastheyshareasplit;abootstrapanalysis
82 Fitzpatricketal.
(A) N. meningitidis
100 E. coli
R. rubrum
N. aromaticivorans
C. crescentus
87
100 B. japonicum
R. palustris
95
87 B. melitensis
B. quintana
95 M. loti
70
A. tumafaciens
100 S. meliloti
R. sphaeroides
100 S. pomeroyi
M. polymorpha
100 R. americana
60 M. jakobiformi
81
R. prowazekii
88 Wolbachia
100 E. ruminantium
0.1 100 A. marginale
(B) 100 B. japonicum (C) 100 B. japonicum (D) N. aromaticivorans
R. palustris R. palustris C. crescentus
91 5494 BB.. qmueilnitteannsais 96 100 AS.. tmuemlialofaticiens 98 60 BM.. qluotiintana
76 60 100 100 ASMC... . mt culroemetlisialcofeatnictiuesns 70 70 100 4090 BBM.. .qm luoetilinitteannsais 76 91100 110000 SBAB.... mjtmauepemlloiilantoefitancicusiimesns
C. crescentus R. palustris
98 100 SR.. psopmhaeeroroyiid es 88 100 R. sphaeroides 100 S. pomeroyi
S. pomeroyi R. sphaeroides
N. aromaticivorans R. rubrum
100 R. americana N. aromaticivorans R. prowazekii
100100 MM.. jpaokloybmifoorrpmhia 88100 NE.. cmoeliningitidis 98100100 WAo. lmbaacrhgiianale
R. rubrum R. rubrum 100 E. ruminantium
FIG.5.—Mainphylogram(A)showstheinferredtopologyoftheproteinconcatenatedalignment;distanceswerederivedfromLogDetTransfor-
mation.Twoofthethreelongbranchesareremovedfromthealignmentandthetreeisredrawn.Thegroupingoftheremaininglongbranchtothefree-
livinga-proteobacteriaisnoted;treesarerootedaroundtheremaininglongbranchforconvenience.(B)ThemitochondriongroupsbesideRhodospirillum
rubrum.(C)TheoutgroupssitbesideR.rubrum.(D)TheRickettsialesgroupbesideR.rubrum.
again revealed weak support for this finding (52%, results optimal supertree to the size of each individual gene tree,
not shown). It should be pointed out that the topologies 77%ofthetopologiesareeitheridenticaltothesupertreeor,
forthefullandshortenedalignmentareidentical;bootstrap wherethetopologiesaredifferent,thisdifferenceisnotsig-
support values for the shorter alignments are lower, how- nificant(i.e.,accordingtotheSHtest,thesedifferencesare
ever. As we systematically removed categories of fast- nogreaterthanwemightexpectfromsamplingeffects,P,
evolving sites from the alignment, the bootstrap support 0.05).Thisiscompatiblewiththehypothesisthatforthese
values decreased across the tree but the overall branching data setsthetopologyinducedbythesupertreeinfigure1
structure never changed. can adequately describe their evolution. Conversely, this
finding is incompatible with a hypothesis that HGT has
Discussion
completely randomized genome evolution. The remaining
To investigate the sister group relationship between 23%oftreesthatshowasignificantdifferencecouldbedo-
thea-proteobacteriaandthemitochondria,itisimperative ingsoasaresultofsystematicbiasesinreconstructingphy-
thatthereisameaningfula-proteobacterialspeciesphylog- logenetic relationships for individual gene trees, HGT, or
eny. We use a supertree-based approach for simultaneous hiddenparalogy.Itisunlikely,though,thatHGTaccounts
determination of species-level phylogeny and analysis of for all 23% of the genes that are incompatible, given the
HGTfor16a-proteobacteriaderivedfrom418genetrees. difficulty of inferring orthology with absolute certainty.
Thebestsupertreereceivesabetterscore(33,196)thanany We have shown that a significantly (v2 test, P , 0.001)
tree from the randomization procedure (YAPTP) which higher percentage of operational genes disagree signifi-
meanstheindividualgenetreeshaveasignificantlyhigher cantly with the supertree, a finding in agreement with
levelofagreementwitheachotherthanwouldbeexpected the complexity hypothesis. A bootstrap analysis of the in-
by chance. The score distributions from real and idealized put orthologous trees also lends support to a robust a-
gene trees (as judged by skewness values) are remarkably proteobacteria phylogeny as the majority of splits are
similar.Itrequiresanaverageof1.6SPRbranchswapsper stronglysupported.Takentogether,wesuggestthatthea-
tree to give ideal data a similar score to the real data, proteobacteria have a meaningful phylogeny.
whereasitrequires5.0SPRswapspertreetoconvertideal Recently,R.rubrumhasbeensuggestedtobethesis-
data into random data. We find that when we prune the ter group of the mitochondria based on a concatenated
GenomePhylogeniesIndicateaMeaningfula-ProteobacterialPhylogeny 83
fast-evolving sites were present in the alignment. As the
fast-evolving sites were removed and the alignments be-
came shorter with a higher proportion of invariant and
slow-evolving sites, we lost phylogenetic signal.
Further evidence that the mitochondria share a com-
mon ancestor with the Rickettsiales to the exclusion of
the other taxa comes from individual mitochondrial pro-
teins. All 15 individual trees indicate a sister group rela-
tionship between the mitochondria and at least one of
the bacteria from this order. It has been suggested that
phylogenetic inferences based on highly conserved mito-
chondrial proteins such as cytochrome oxidase subunits
cox3 have been suggested as better indicators of relation-
ships than fast-evolving mitochondrial genes that may
infer incorrect relationships as an artifact of long-branch
attraction (Lang, Gray, and Burger 1999). Taking the in-
ferred evolutionary history of cox3 alone, we can see
FIG.6.—Thisphylogramdepictsareducedconstrainedtree.Thistree that this gene would suggest that the mitochondria and
onlycontainstheshort-branchedtaxa.Long-branchedtaxaarerandomly Rickettsialesareeachothers’closestrelatives(table1)with
attachedtotheconstrainedtree.Relativebranchlengthsforthenewen-
bootstrap support for this relationship being relatively
largedconstrainedtreeweredeterminedinPHYMLusingtheoriginalcon-
catenatedalignment.Thisprocedurewasrepeated100timestogive100 strong (;80%).
randomconstrainedtrees. Abetterunderstandingoftheorigin andevolution of
themitochondrionwillonlybepossiblewhenamorerep-
alignmentof31mitochondrion-encodedgenes(Esseretal. resentative sample of the a-proteobacteria is fully se-
2004). Our analysis utilized 15 mitochondrial genes that quenced. It is entirely possible that there are many more
haveaphylogenythatisnotsignificantlydifferent(SHtest, mitochondrion-like bacterial species yet to be identified.
P,0.05)totheproposed supertree. Using thisapproach, Currently, only 0.4% of all existing bacterial species have
wehaveminimizedtheobfuscatinginfluenceofincluding beenidentifiedandformallydescribedletalonesequenced.
genes whose history might not be compatible with one Anotherconsiderationinsuchananalysismustbethedraw-
another. Our data point to the conventional sister group backs associated with the use of eukaryote mitochondrial
relationship of Rickettsiales and the mitochondrion. For DNAandthehighrateofnucleotidesubstitutionassociated
completeness,weperformedouranalysesonalargeralign- withthesegenes.Theseproblemscanbeexacerbatedwhen
ment containing all 28 mitochondrion-encoded proteins, we use mitochondrion sequences in studies with bacterial
and the conclusions of these analyses agree with the find- homologs that have divergence times of approximately
ings of our initial analysis (results not shown). When we 1.5–2billionyears(Sicheritz-Ponten,Kurland,andAndersson
reexamined the 31-gene sequence alignment of Esser 1998). The end result of such an analysis can be severely
et al. (2004), we found alternative topologies depending affected by long-branch attraction. Throughout this analy-
onwhatmethodwasutilizedtoremovefast-evolvingsites. sis,weweremindfuloftheseproblemsandendeavoredto
Whenfast-evolvingsiteswerestrippedinthesamemanner take steps such as the use of LogDet distance matrices to
as their analysis, we recreated identical results to those helppreventmisleadingresults,andweinvestigatedthepo-
authors (fig. 1A). Alternatively, when fast-evolving sites sition of long-branched taxa in the absence of other long
categorized by a gamma distribution are removed, the branches. Another approach that is analogous to a trans-
RickettsialesandnotR.rubrumareinferredtobethesister version analysis of nucleotide sequences is to recode the
grouptothemitochondrion(fig.1B).Theseresultsraisecon- inputalignmentintothesixDayhoffcategoriesandsubse-
cerns regarding the method used to strip out fast-evolving quentlyinfertheirphylogeny.Thisapproachhastheadvan-
sites. The method described by Hansmann and Martin tage that we could use a Bayesian approach and optimize
(2000) strips sites out based on the number of character ageneraltime-reversiblesubstitutionmatrix,ratherthanas-
states found at a particular site, while the discrete gamma suming an ad hoc amino acid substitution matrix. We in-
methodusesMLestimationtoplacesitesintodifferentsite vestigated such an approach on the 15-gene alignment
categories.Clearly,thesemethodscategorizesitesverydif- (gapped positions removed) and found that the grouping
ferently,andthemeritsofeachwillneedfurtherstudy.Inour of the mitochondrion and Rickettsiales are again strongly
analysis, wehavemoreextensive taxon sampling, and also supported.
we erred on the side of caution and removed fast-evolving Esseretal.(2004)makethepointthattheoverallfer-
sites using both methods described above until we found mentativephysiologyofR.rubrumisquitesimilartoeukar-
alignments that passed a v2 test of amino acid composi- yotes that lack mitochondria or contain anaerobic
tionalheterogeneity.Phylogeneticinferencesofbothresul- mitochondria. Similarly, it can also be argued that there
tant alignments yielded similar conclusions in that the is a strong affinity between mitochondria and Rickettsia
Ricketsialesareinferredtobethesistertaxonofthemito- for the genes coding for components of the Krebs cycle
chondrionwithvaryingdegreesofsupport(65%and52% (Anderssonetal.1998),therespiratorychain(Gray,Burger,
bootstrap support, respectively). Higher levels of support and Lang 1999), and the translation system (Sicheritz-
for these inferences were observed when categories of Ponten, Kurland, and Andersson 1998).