微生物生态的多元分析.pdf

《微生物生态的多元分析.pdf》由会员分享，可在线阅读，更多相关《微生物生态的多元分析.pdf（19页珍藏版）》请在淘文阁 - 分享文档赚钱的网站上搜索。

1、M IN I R E V I E WMultivariateanalysesinmicrobialecologyAlban RametteMicrobial habitat group,Max Planck Institute for Marine Microbiology,Bremen,GermanyOnlineOpen:This article is available free online at www.blackwell-Correspondence:Alban Ramette,Microbialhabitat group,Max Planck Institute forMarine

2、 Microbiology,Celsiusstrasse 1,28359 Bremen,Germany.Tel.:149 4212028 863;fax:149 421 2028 690;e-mail:aramettempi-bremen.deReceived 17 January 2007;revised 18 July2007;accepted 20 July 2007.First published online 25 September 2007.DOI:10.1111/j.1574-6941.2007.00375.xEditor:Ian HeadKeywordsordination;

3、multivariate;modeling;statistics;gradient.AbstractEnvironmental microbiology is undergoing a dramatic revolution due to theincreasing accumulation of biological information and contextual environmentalparameters.This will not only enable a better identification of diversity patterns,but will also sh

4、ed more light on the associated environmental conditions,spatiallocations,and seasonal fluctuations,which could explain such patterns.Complexecological questions may now be addressed using multivariate statistical analyses,which represent a vast potential of techniques that are still underexploited.

5、Here,well-established exploratory and hypothesis-driven approaches are reviewed,so asto foster their addition to the microbial ecologist toolbox.Because such tools aimat reducing data set complexity,at identifying major patterns and putative causalfactors,they will certainly find many applications i

6、n microbial ecology.IntroductionMicrobial ecology is undergoing a profound change becausestructurefunction relationships between communities andtheir environment are starting to be investigated at the field,regional,and even continental scales(e.g.Hughes Martinyet al.,2006;Ramette&Tiedje,2007a,b).Be

7、cause DNAsequences are being accumulated at an unprecedented ratedue to high-throughput technologies such as pyrosequen-cing(Edwards et al.,2006a,b),single-cell genome sequen-cing(Zhang et al.,2006),or metagenomics(Venter et al.,2004;Field et al.,2006;Gill et al.,2006),future challengeswill very lik

8、ely consist of interpreting the observed diversitypatterns as a function of contextual environmental para-meters.This would help answer fundamental questions inmicrobial ecology such as whether microbial diversityresponds qualitatively and quantitatively to the same factorsas macroorganism diversity

9、(Horner-Devine et al.,2004;vander Gast et al.,2005;Green&Bohannan,2006;HughesMartiny et al.,2006).Most obstacles encountered by microbial ecologistswhen they try to summarize and further explore large datasets concern the choice of the adequate numerical tools tofurtherevaluate the data statisticall

10、y and visually.Such tools,which have been developed by community ecologists towork on distribution and diversity patterns of plants andanimals,could be readily applied in microbial ecology.Although multivariate analyses of community diversitypatterns are well described in the literature,microbialeco

11、logists have used multivariate analyses either rarely ormostly for exploratory purposes.A brief survey of theliterature confirms this trend(Table 1;Fig.1).Table 1indicates that bacterial studies rank third after plant andfish studies for their use of multivariate analyses.Complexdata sets are mostly

12、 explored via principal componentanalysis,or cluster analysis,and hypothesis-driven techni-ques such as redundancy analysis,canonical correspondenceanalysis(CCA),or Mantel tests aremore rarely used(Fig.1).Axis 1(horizontal)clearly differentiates microscopic(bac-teria,microorganisms,fungi)from macros

13、copic(fish,bird,plant,insect)life,and this may be related to the use ofmore exploratory methods(e.g.cluster analysis,PCA)inthe first group.It is important to state that the figurespresented in Table 1 and Fig.1 have to be taken withcaution because many articles do not include a descriptionof statist

14、ical approaches in their titles or abstracts,and soFEMS Microbiol Ecol 62(2007)142160c?2007 Max Planck SocietyJournal compilationc?2007 Federation of European Microbiological SocietiesPublished by Blackwell Publishing Ltd.the table is certainly biased and incomplete.However,thepoint of the table was

15、 both to identify some general trendsin the literature and to give one example of the usefulness ofmultivariate analysis to analyze a data table.This review aims atpresentingsomecommon multivariatetechniques in order to foster their integration into themicrobial ecologists toolbox.Indeed,it is no lo

16、nger possibleto gain a full understanding of Ecology and Systematicswithout some knowledge of multivariate analysis.Or,contra-riwise,misunderstanding of the methods can inhibit ad-vancementofthescience(James&McCulloch,1990).Suchareview is ambitious because it tries to provide a few guide-lines for a

17、 very vast discipline that is still under development.For this reason,it cannot be exhaustive and does not pretendto offer in-depth coverage of all selected topics.The review islargely inspired by descriptions,comments,and suggestionsoriginating from multiple,highly recommended sources(terBraak&Pren

18、tice,1988;James&McCulloch,1990;Legendre&Legendre,1998;Leps&Smilauer,1999;ter Braak&Smilauer,2002;Palmer,2006),where detailed informationabout each technique can be obtained.In the first part,data type and preparation arereviewed asa necessary basis for subsequent multivariate analyses.Second,common

19、multivariate methods(i.e.cluster analysis,principal component analysis,correspondence analysis,multidimensional scaling)and a few statistical methods totest for significant differences between groups or clusters aredescribed,focusing on the methods main objectives,appli-cations,and limitations.Beyon

20、d the mere identification ofdiversity patterns,microbial ecologists may wish to correlateor explain those patterns using measured environmentalparameters,and this approach is addressed in the third part.Special emphasis is placed on a few methods that haveproven useful in ecological studies,namely r

21、edundancyanalysis,CCA,linear discriminant analysis,as well as varia-tion partitioning.The final part provides practical consid-erations to help researchers avoid pitfalls and choose themost appropriate methods.1.01.01.5clusterPCAMDSPCoACCARDAMANOVAMantelANOSIMCVABacteriaPlantFungBirdInsect001.0FishM

22、icrobFig.1.Correspondence analysis of method usage in various scientificfields.In this symmetrical scaling of CA scores,the first two axesexplained 47.3%and 35.8%of the total inertia of Table 1,respectively.The gray areas were drawn to facilitate the interpretation.Complete rownames(scientific field

23、s;full circles)and column names(methods;whitetriangles)are given in Table 1.Methods(triangles)located close to eachother correspond to methods often occurring together in studies.Thedistance between a scientific field point and a method point approx-imates the probability of method usage in the fiel

24、d.Table 1.Usage(%)of multivariate methods in different fieldsKeywordswExploratory analysisHypothesis-driven analysisTotal numberzClusterPCAMDSPCoACCARDAMANOVAMantelANOSIMCVABacter?48.5384.50.43.21.81.30.40.91.11141Microb?45.840.23.91.12.22.21.11.70.61.1179Plant?40.328.54.61.715.53.71.92.30.60.93335F

25、ung?5427.22.81.18.52.80.91.10.21.4563Fish?30.133.79.80.313.52.73.62.92.31.21464Bird?4120.55.40.721.23.52.14.20.50.9429Insect?54.313.76.10.811.54.43.531.11.7637A literature search was performed with the Thomson ISI research tool with the following parameters(Doc type,all document types;language,allla

26、nguages;databases,SCI-EXPANDED,SSCI,A&HCI;Timespan,19002006)on December 13,2006 in the titles and abstracts of the articles only.wAsterisks were placed at the end of each keyword to accommodate for variations.Each keyword was additionally combined with the followingtechnical designations:cluster,clu

27、ster analysis;PCA,principal component analysis;MDS,multidimensional scaling;PcoA,principal coordinate analysis;CCA,canonical correspondence analysis;RDA,redundancy analysis;Mantel,Mantel test,or CVA,canonical variate analysis.zTotal number refers to the total number of publications identified by eac

28、h keyword and all its combinations.The ordination based on correspondenceanalysis of the raw number is depicted in Fig.1.FEMS Microbiol Ecol 62(2007)142160c?2007 Max Planck SocietyJournal compilationc?2007 Federation of European Microbiological SocietiesPublished by Blackwell Publishing Ltd.143Multi

29、variate analyses in microbial ecologyData types and data preparationData setsThe initial multivariate data set may consist of a table ofobjects(e.g.samples,sites,time periods)in rows andmeasured variables for those objects in columns.This tablestructure is the standard used in the present review.Whe

30、nthe latter variables are biological taxa,the columns willsimply be designated as species thereafter.It is critical toclearly identify what corresponds to objects and variables inthe data set.Indeed,objects in one study may be species oroperational taxonomic units(OTU)for which catabolicprofiles,gen

31、e presence or polymorphism,etc.are measured.In another study where samples from different sites arecompared based on,for instance,community fingerprintingtechniques,objects can now be samples and species vari-ables.This distinction is important because procedures thatanalyze relationships among obje

32、cts or among variables aredifferent.Objects are defined a priori by the samplingstrategy before making observations and variable measure-ments.Besides,most multivariate analyses assume indepen-dence between objects(or samples),i.e.observations madeon an object are not a priori dependent on those mad

33、e onanother object.Variables,however,can be found to beintercorrelated to various degrees,but this is not necessarilyknown in advance.Initial data sets can also consist ofdistance matrices where pairwise dissimilarities betweenobjects are calculated.The original table of raw data is notalways availa

34、ble,e.g.for DNADNA hybridization values,phylogenetic distances,and thus specific multivariate tech-niques have to be considered to deal with data matrices.Data transformationsIn multivariate data tables,measured variables can bebinary,quantitative,qualitative,rank-ordered,classes,fre-quencies,or eve

35、n a mixture of those types.If variables donot have a uniform scale(e.g.environmental parametersmeasured in different units or scales)or an adequate format,variables have to be transformed before performing furtheranalyses.Each qualitative variable has to be recoded as a setof numerical variables tha

36、t replace it in the numericalcalculations.One way to do so is to create a series ofdummy variables that correspond to all the states of thequalitative variable.For instance,if the variable season hasto be recoded,four associated variables will be constructed,and for each object the value 1 will be g

37、iven to thecorresponding season when it occurs,and 0 for the threeother seasons when it is absent.Many statistical packagesautomatically perform this recoding.Standardization provides dimensionless variables andremoves the undue influence of magnitude differencesbetween scales or units.A common proc

38、edure is to applythe z-score transformation to the values of eachvariable.Foreach variable,it consists of(1)computing the differencebetween the original value and the mean of the variable(i.e.centering)and of(2)dividing this difference by the SD ofthe variable.Normalizing transformations aim at corr

39、ecting the dis-tribution shapes of certain variables,which depart fromnormality.One thus tries to obtain a homogeneous var-iances for variables,conditions under which multivariateprocedures often perform better.Different mathematicaltransformations can be used to normalize the x values of avariable:

40、for instance,the arcsin(px)transformation can beapplied to percentages or proportions,log(x1c)to variablesdeparting strongly from a normal distribution,andp(x1c)to less problematic cases,with c being a constant that isadded to avoid mathematically undefined computations.The c constant is generally c

41、hosen so that the smallestnonzero value is obtained by computing x1c in the formerfunctions.The constant should also be of the same order ofmagnitude as the variable(Legendre&Legendre,1998).To make community composition(either presenceabsence or abundance)data containing many zeros suitablefor analy

42、sis by linear methods such as principal componentanalysis(PCA)or canonical redundancy analysis(RDA),theHellinger transformation Eq.(1)is one of five transforma-tions that give good results(Legendre&Gallagher,2001).The chord transformation is a useful transformation thatalso gives less weight to rare

43、 species in the species table Eq.(2).The transformations are given byy0ijffiffiffiffiffiffiyijyir1y0ijyijffiffiffiffiffiffiffiffiffiffiPpj1y2ijs2where yijis the original species value for site i and species j,yi1represents the sum of all species values for site i(i.e.sumper row),p is the number of s

44、pecies in the table(number ofcolumns),and y0ijrepresents the resulting,transformed spe-cies value(Legendre&Gallagher,2001).These transforma-tions are particularly recommended when rare species arenot truly rare,i.e.when they mostly occur because thesampling was performed blindly,as generally done in

45、 soilor marine microbial ecology.Further data transformationscan be found in Sokal&Rohlf(1995)and Legendre&Legendre(1998).The way to deal with missing data is a discipline on itsown(Legendre&Legendre,1998).Briefly,one can eitherdelete rows or columns containing the missing value(s),ortry to replace

46、the missing values by mathematical estimatesinferred from values obtained from other objects in the dataFEMS Microbiol Ecol 62(2007)142160c?2007 Max Planck SocietyJournal compilationc?2007 Federation of European Microbiological SocietiesPublished by Blackwell Publishing Ltd.144A.Rametteset.Inthe lat

47、tercase,it is still difficult to provide ecologicallymeaningful explanations for these estimates.In any case,thespecific handling of missing data should be reported by theinvestigator.When dealing with matrices,it is possible to change asimilarity matrix(S)into a dissimilarity matrix(D)byapplyingthe

48、followingtransformations:D=1?S,D=p(1?S),or D=p(1?S2).To normalize any D matrixto the interval 01,one can compute D/Dmax,or(D?Dmin)/(Dmax?Dmin),where Dmaxand Dminrepresentthe highest and lowest values of D,respectively(Legendre&Legendre,1998).Exploratory analysesVisualization and exploration of compl

49、ex datasetsThe basic aim of ordination and cluster analysis is torepresent the(dis)similarity between objects(e.g.samples,sites)based on values of multiple variables(columns)associated with them,so that similar objects are depictednear from each other and dissimilar objects are foundfurther apart fr

50、om each other.Exploratory multivariateanalyses are thus useful to reveal patterns in large data sets,but they do not directly explain why those patterns exist.This latter point is addressed in the third part of the review.Cluster analysis and association coefficientsCluster analysis encompasses seve