• Aucun résultat trouvé

Exploratory analysis of cancer SAGE data

N/A
N/A
Protected

Academic year: 2021

Partager "Exploratory analysis of cancer SAGE data"

Copied!
7
0
0

Texte intégral

(1)

HAL Id: hal-00363020

https://hal.archives-ouvertes.fr/hal-00363020

Submitted on 26 Apr 2010

HAL is a multi-disciplinary open access

archive for the deposit and dissemination of

sci-entific research documents, whether they are

pub-lished or not. The documents may come from

teaching and research institutions in France or

abroad, or from public or private research centers.

L’archive ouverte pluridisciplinaire HAL, est

destinée au dépôt et à la diffusion de documents

scientifiques de niveau recherche, publiés ou non,

émanant des établissements d’enseignement et de

recherche français ou étrangers, des laboratoires

publics ou privés.

Exploratory analysis of cancer SAGE data

Ricardo Martinez, Richard Christen, Claude Pasquier, Nicolas Pasquier

To cite this version:

Ricardo Martinez, Richard Christen, Claude Pasquier, Nicolas Pasquier. Exploratory analysis of

cancer SAGE data. PKDD’2005 International Conference, Oct 2005, Porto, Portugal. 6 p.

�hal-00363020�

(2)

Ri ardoMartinez

1

,Ri hardChristen

2

,ClaudePasquier

2

,Ni olasPasquier

1

1

LaboratoireI3S(CNRSUMR-6070),UniversitédeNi e, 2000routedeslu ioles,LesAlgorithmes,06903Sophia-Antipolis,Fran e;

{rmartine,pasquier}i3s.uni e.fr

2

LaboratoiredeBiologieVirtuelle(CNRSUMR-6543),UniversitédeNi e, CentredeBio himie,Par Valrose,06108 Ni e edex2,Fran e;

{ hristen, laude.pasquier}uni e.fr

Abstra t. Usingseveral analysete hniquesfor the hierar hi al lustering ofaSAGEexpressiondatasetof822tagsfrom74tissuesamples(normaland an er)weshowthat leaningthedataset(tagsandexperiments)is riti al andthat attributionofa tagto ageneisnoteasy.Comparisonof an ers fromvarioustissuesisadi ulttaskastissuesamples lustera ording to tissueoriginandnotas an erornormal.

1 Introdu tion

TheSAGEmethodisbasedonthesequen ingof on atemersofshort(14basepairs; re ently 17 bp) sequen e tagsthat originatefrom the 3'-nearest utting site of a restri tion enzyme) to estimatetrans ripts abundan e[VZVK95℄, to estimate the expressionlevelofeukaryoti trans riptswithoutpriorknowledgeoftheirsequen es and is moresensitivethan the EST method [SZL+04℄, but requires knowledgeof the ompletegenome.TheadvantageoftheSAGEmethodistoperformarandom samplingof trans riptsin aparti ulartissue,withlittlesequen ingeort.

Thedatasetproposedforanalysis omprisesseveraldi ulties:

1. PCRandsequen ingmayprodu eanumberoferrors.Asingleerrormayleadto nonre ognitionofatrans riptorwrongattribution.Sometagsmaybepresent in more than one gene. Finally, sin e restri tion enzymes may not ut with 100%e ien y,sometagsmaybewrong.

2. Tissuesamplesoriginatefrom twodierentsour es(i.e.bulkor ellline) that may inuen e gene expression. Can erous tissue are usually provided after surgery,a an ersamplemay ontainmorehealthytissuethan an er,leading toawrong identi ation.

3. Analyzes using DNA hips on luded that an er ellsare more alike normal ellsofthesametissuethan an er ellsfromadierenttissue:therearemany moretissue-spe i genes than genes involved in an ers[RSE+00,SRW+00℄. Thus,tryingto lassifyintwo lasses,normalversus an er,inordertoidentify spe i tags an bedi ult. Also, an ersmayhavedierentorigins (deregu-lationof on ogenesversusbreakdownsof hromosomesforexample) sear hing fortwo lassesonlymaybeproblemati .

4. Interpretations. Even after removal of tags that do not show any signi ant hange among samples, manytags remain to be lassied. Onemaythen use

(3)

exampleGeneOntology,ifea htagislinkedtoagene.

Themaingoalofouranalysiswastoinvestigatetheinuen eof leaningthedataset. We propose to validate removal ofspurious tags or experimentsand therefore in- reasethesignal.Inanexploratoryanalysisweusedthesmalldataset.Thispaper fo usesonthefollowingsteps:i)Pruningofnon-signi anttags;ii)data normaliza-tion;iii) sele tionof dierentiallyexpressed genes;iv)deletionofoutlier biologi al onditions;v) lassi ationofbiologi al onditions.

2 Tags sele tion

TagsareoftenannotatedbasedontheSAGEGenieprin iples[BOG+02℄andlinked to a series of expression data (often EST sequen es), a step that is di ult to automate. It isoften di ult to understand and appre iate themethods usedfor tag attribution, we therefore developed spe i tools. First, every human ENST sequen e was downloaded from Ensembl. Tags present in trans ripts of a single gene were labelledas good(436) attributed orrespondingENSG numbers

3 . Tags presentin trans riptsoriginatingfromseveralgeneswerelabeledasbad(219)and removedfromfurtheranalysis.

Next,allEMBLhumansequen es(in ludingESTs)weredownloadedtosear hnon attributedtags(167).Everysequen ere ognizedwasblastedforENSGattribution. This stepled to afurther 80tags attributed toa ENSG number.Reasonsfor tag nonattribution arelikelytobe:i)lo ationinaregionnotyetidentiedas agene; ii)lo ationinthemito hondrialgenome(veryfewprotein odinggenes),whi hwas nottakenintoa ount;iii)tagresultingfromthepartial digestionofatrans ript, andthereforenotlo atedinthe3-mostdomain.

Atthispointwehad learlylesstagslinkedtogenesthanifwehadusedatoolsu h asSAGEGenie.Butthersttagofthelistwaslinkedtoamito hondrialsequen e by SAGE Genie, while at the Global Gene ExpressionGroup proje t it mapped toUnigeneHs.476965(G1/Stransition ontrolprotein-bindingproteinIEF-8502)

4 . The SAGE Genie linked this tag to a sequen e of a ession number BE874599. Blast ofthis sequen eprovided ahiton themito hondrial humangenome,but at a position that was identied as `16S ribosomal sequen e'. Su h sequen e hasno polyAtailofanysort,anddoesnot ontainarepeatofAanywhereinthesequen e. At this step we are rather ondent that every data resulting from large s ale analysisusingwebbasedtools,shouldbe riti allyassessedeitherusingtwodierent publi toolsor ad-ho s riptsanddatabases

5 .

3 Algorithms and methods

Weused theSigni an e Analysisof Mi roarrays(SAM) 6

method to sele t dier-entiallyexpressed genes.SAM omputesastatisti

di

foreverygene

i

, measuring

3

http://www.ensembl.org/ 4

http://s ien epark.mdanderson. org/ ggeg 5

http://www.n bi.nlm.nih.gov/ 6

(4)

( an er bulk, an er ell line, normal bulk and normal ell line). The uto for signi an eDelta was xedat0.21implyingaFalseDis overyRateof5%. Taking into a ount onditionvariations and in parti ular outliers that introdu e noise in the lassi ationis riti al [LMV04℄. Thus, we developed amethodology for nding outliers using Prin ipal Component Analysis (PCA) and hierar hi al lusteringmethods:

1. UsingPCAasanexploratorytooltodeterminetheoptimalnumberof lusters. 2. Applyinghierar hi al lusteringalgorithmstoidentifyoutliersandremovethem. 3. Applying again PCA analysis to verify that variability level is not de reased

whenea hofthese onditionsisremoved.

4. Clustertoverifythatthe lusteringwasimproved.

Wetested 5algorithms(K Means,Fanny, PartialLeastSquares,UnweightedPair GroupsMethod Average(UPGMA) andDIvisiveANAlysis(DIANA)) and5 mea-sures of distan e (Eu lidean, Pearson, Manhattan, Spearman and Tau) a ording to3dierent onsisten ymeasures(averageproportionofnon-overlap,average dis-tan e between lusters and average distan e between luster means) [DD03℄. We sele ted UPGMA and DIANA algorithms and Pearson, Eu lidean and Spearman distan esthat arethemoste ientwiththisdataset.

4 Experimental results

4.1 Biologi al ondition sele tion

The7pan reas onditionsaredistributedin3 lasses: an er ellline(C1Ce,C2Ce andC3Ce),normal ellline(N1CeandN2Ce)andnormalbulk(N3BuandN4Bu), asisshownbytherst3PCA omponentsthatexplain98.59%ofthetotalvarian e. The hierar hi al trees obtained for the dierent distan e measures are shown in gure1(a).TreesobtainedwiththeUPGMAandtheDIANAalgorithmsaresimilar.

(a)All onditions

(b)Withoutoutliers

Fig.1.Hierar hi al lusteringofthepan reas onditions.

Using Pearson and Eu lidean distan es, ondition Pan reasC3Ce is pla ed in an isolated luster,andwhentheSpearmanmeasureisuseditisasso iatedwithnormal onditions.Removingthis ondition,therst3 omponentsexplain99.03%of the varian eandtheresultof lusteringisshowningure1(b).

(5)

allexperimentswe anseethatthevarian eexplainedbytherst3 omponentsis alwaysimprovedwhenoutlier onditionsareremoved.

Organ/Tissue PCA Outliers PCAwithoutOutliers

(rst3 omp.) (rst3 omp.)

Brain 98.46% {N4Ce,C1Bu,C14Bu,C5Bu,C9Ce} 99.02%

Breast 95.57% {C6Bu} 97.38%

Colon 98.56% {} 98.56%

Ovary 93.60% {N1Ce,N2Ce,C4Ce,C6Bu} 97.91% Prostate 98.02% {N1Bu,C7Bu,C9Bu, C8Ce,C1Ce} 98.70%

Pan reas 98.59% {C3Ce} 99.03%

Table1.PCAanalysis of onditionsbytissues.

Forea htissue,weobservethree lasses: an er,bulkand ellline,withbulkand ell line learlyseparated(seegure1(b)).Thisobservationtherefore onrmsprevious analyzesthatshowed ellsour etobeof ru ialinuen eongeneexpression.

4.2 Hierar hi al lustering of biologi al onditions

WeappliedPCAanalysisandfoundthattherst6 omponentsexplain98.22%of the varian e orresponding to the 6tissue lusters. Comparingthese resultswith PCAanalysisontheinitialdatasetshowedthatgeneand onditionsele tionshave eliminateddatanoise.

Then, weappliedtheUPGMAandDIANA algorithmsto the leaned datasetand the tree obtainedby onsensus for both algorithms, and for the Pearsonand the Spearmandistan es, isshownin gure2.For theEu lideandistan e, the distribu-tionissimilarbutbran hesto theleavesarelonger.

Comparing lustering trees obtainedwith the initialdataset (not shown) and g-ure 2 learly showedthat thesele tionpro essimproved dataquality sin elength ofterminalbran heswere onsiderablyredu ed.We anobservearstdegree las-si ationbytissuethatisa urateforPan reas,Brain,Breast,ColonandProstate tissues,butmixesOvarytissue onditionswithothertissue onditions.We analso see a lear se ond degree lassi ation, among onditions of the same tissue, by ellsour e:bulkand ellline.AmongPan reas,Breast,BrainandColon ondition lusters, we anobserveathirddegree lassi ationbystate: an erandnormal.

Fig.2.Hierar hi al lusteringof onditions.

(6)

re-ignored.Webelievethatwhen ondu tingstudiesforndinginterestinggene an- er knowledge involvingmultiple tissuesSAGE libraries, thestudy must berst orientedtowardade ompositionofthe onditionsbytissuesandthenby ellsour es to nallyfo ustheanalysison ellstates.

Eventually,weappliedtheC5.0unsupervised lassi ationmethodtoprodu e las-si ation rules of biologi al onditions by tissue, ell state and ell type. Three dierent lass attributes hara terizing ea h ondition were reated: tissue type (Pan reas,Ovary, Brain, Prostate and Breast), ell sour e (bulk or ell line) and ellstate( an erandnormal).Boostingand rossvalidationoptionswerea tivated. Thenumbersofruleswithmaximala ura ygeneratedforea h lassde omposition of onditionsareshownintable 2.

Class NumberofrulesMaxa ura y

Bulk 5 100%

Cellline 5 100%

Can er 1 80%

Normal 3 80%

All6tissuestypes 1 60% Table 2.Rulesby lassandtheirmaximala ura y.

Using ellsour e lassi ation,5exa ts rules, i.e. with 100% a ura y, were gen-erated. For ell state lassi ation,only 1and 3 rules respe tively, all with with only80% ofa ura y,weregenerated.Consideringtissue lassi ation,only1rule with 60% a ura y was generated. This result is logi al sin e there are 6 dier-ent tissues, thus disturbing the lassi ation, and ells from dierent tissues but originatingfrom elllinestendtobe omemoresimilarfromthetagexpression lev-els viewpoint. These results onrmthat in the small leaned dataset,there isan intrinsi divisionof onditionsby ellsour ethatismorenaturalthanby ellstate.

5 Con lusion

Most SAGE studies madeuse tagsof 14 bp.However, are entstudy showed the learadvantageofusingatagof15bp[DBB+05℄.Evenlongertagswillbebetter. Re ently, the SAGE proto ol was enhan ed with anew tagging enzyme(MmeI), whi h produ es 21-22 bases tags [SSR+02℄, allowing dire t mappingto the tran-s ripts[VC04℄.Whennumeroustagsareavailableremovingtagspresentonlyon e, that may result from errors, is possible. Sequen e errors havelittle ee t on the quanti ation of moderately expressed genes but not for rare trans ripts. About 6.7% ofLongSAGEditagswill havea quiredmutations priortoligation, loning andsequen ing[VC04℄,arguingforarobusttagattributionto atrans ript. Only reliably annotated tags an be in luded in the nal analysis [SSL+04℄. An-notation ofSAGEtagsto genes andtheir orrespondingUnigene luster numbers revealedthat onaverageonly30%ofalltags(in ludinglessabundanttags) ould bereliablyannotatedbasedontheSAGEGenie prin iples[BOG+02℄.Annotation improvedto about 70% for tagswith intermediateto abundantexpression levels. Remainingtagseither ouldnotreliablybeasso iatedwithagene(e.g. annotated

(7)

results[DBB+05℄and using asingle omputerprogram andasingle sour eof se-quen e data (annotations) would resultin a weakeranalysis.Wehavealso shown in oheren e of resultsbetween dierentpubli web tools,and an obvious errorof gene attribution for the rst tag at least. Removing outlier experiments also de- reases noiseand in reasesreliability of lustering.Finally, wesawthat sear hing for lassi ation rules identifying normal and an erous tissues among tissues of dierentoriginsisdi ultasrulesofmaximala ura ydis riminatetissueorigins. Usingseveraldatasets ontainingea honenumeroussamplesfromthesametissue ouldimprovetheresults.

Referen es

[BOG+02℄ BoonK.,OsorioE.C.,GreenhutS.F.,S haeferC.F.,ShoemakerJ.,PolyakK., MorinP.J.,BuetowK.H.,StrausbergR.L.,DeSouzaS.J.andRigginsG.J. Ananatomy ofnormalandmalignantgeneexpression.Natl.A ad.S i.USA,99(17):11287-292,2002. [DD03℄ Datta Su. and Datta So. Comparisons and validation of statisti al lustering

te hniquesformi roarraygeneexpressiondata. Bioinformati s,19(4):459-466,2003. [DBB+05℄ DinelS., Boldu C., BelleauP.,BoivinA., YoshiokaM.,CalvoE.,Piedboeuf

B.,SnyderE.E.,LabrieF.andSt-AmandJ. Reprodu ibility,bioinformati analysisand powerof theSAGEmethodtoevaluate hangesintrans riptome. Nu lei A ids Res., 33(3):e26,2005.

[LMV04℄ Loguinov A.V., Mian I.S. and Vulpe C.D. Exploratory dierential gene ex-pressionanalysis inmi roarray experimentswithnoorlimited repli ation. Gen.Bio., 5(3):R18,2004.

[NSS01℄ NgR.T.,SanderJ.andSleumerM.C. Hierar hi al lusteranalysisofSAGEdata for an erproling. Pro .BIOKDD onf.,pp65-72,2001.

[PGJC04℄ Pasquier C., Girardot F., Jevardat deFombelle K. and Christen R. THEA: ontology-drivenanalysisofmi roarraydata Bioinformati s,20(16):2636-2643, 2004. [RSE+00℄ RossD.T.,S herfU.,EisenM.B.,PerouC.M.,ReesC.,SpellmanP.,IyerV.,

JereyS.S.,VandeRijnM.,WalthamM.,Pergamens hikovA.,LeeJ.C.,LashkariD., ShalonD.,MyersT.G.,WeinsteinJ.N.,BotsteinD.andBrownP.O.Systemati variation ingeneexpressionpatternsinhuman an er elllines.Nat. Genet.,24(3):227-35,2000. [SSR+02℄ SahaS.,SparksA.B., RagoC., AkmaevV.,WangC.J.,VogelsteinB., Kinzler K.W. and Vel ules u V.E. Using the trans riptome to annotate the genome. Nat. Biote hnol.,20(5):508-512,2002.

[SRW+00℄ S herf U., Ross D.T.,Waltham M., Smith L.H., LeeJ.K., Tanabe L., Kohn K.W.,ReinholdW.C.,MyersT.G.,AndrewsD.T.,S udieroD.A.,EisenM.B.,Sausville E.A., Pommier Y., Botstein D., Brown P.O. and Weinstein J.N. A gene expression databaseforthemole ularpharma ologyof an er. Nat.Genet.,24(3):236-44,2000. [SSL+04℄ Shippy R., Sendera T.J., Lo kner R., Palaniappan C., Kaysser-Krani h T.,

WattsG.andAlsobrookJ. Performan eevaluationof ommer ialshort-oligonu leotide mi roarraysand theimpa tof noise inmaking ross-platform orrelations. BMC Ge-nomi s,5(1):61,2004.

[SZL+04℄ SunM.,ZhouG.,LeeS.,ChenJ.,ShiR.Z.andWangS.M. SAGEisfarmore sensitivethanESTfordete tinglow-abundan etrans ripts. BMCGenomi s,5(1),2004. [VZVK95℄ Vel ules uV.E.,ZhangL.,VogelsteinB.andKinzlerK.W. SerialAnalysisof

GeneExpression. S ien e,270:484-487,1995.

[VC04℄ Viat heslav R.A. and Claren e J.W. Corre tion of sequen e-based artifa ts in serialanalysisofgeneexpression. Bioinformati s,20(8):1254-1263,2004.

Figure

Fig. 1. Hierarhial lustering of the panreas onditions.
Fig. 2. Hierarhial lustering of onditions.

Références

Documents relatifs

It is not a problem if the contact person is not totally familiar with the topic 6 , but he/she must indicate the stakeholders having the most diverse opinions possible for each

This text is based on notes (taken by R. Thangadurai) of three courses given at Goa University on December 15 and 16, 2014, on Algebraic independence of Periods of Elliptic

Specically, we will use a) ergodic theory to show that one cannot reliably predict by analogy with the past, even in deterministic systems, chaotic or not, and b) Ramsey theory to

This volume contains the proceedings of the International Conference on Complex Analysis and Related Topics – The 13th Romanian-Finnish Seminar, which was held at the University

In addition, the latency tolerance range (LTR) enable the comparison of results regarding the issue of latency between the different musical instrument groups. The results and

So, the personalized approach to the processing of medical information is characterized by a number of problems, namely: the uncertainty of the data presented, the classifica- tion

The  program of  the intern  is in  a  first step  to investigate some bandwidth allocation algorithms with cyclic scheduling and to investigate the main performance. issues

This is followed by the NEST (preferred NESTip version) for models with increasing number of di- mensions, where the rank of the data eigenvalues among the corresponding eigenvalues