A cut-off based approach for gene expression analysis of formalin-fixed and paraffin-embedded tissue samples

A cut-off based approach for gene expression analysis of formalin-fixed and paraffin-embedded tissue samples
of 8
All materials on our website are shared by users. If you have any questions about copyright issues, please report us to resolve them. We are always happy to assist you.
Related Documents
  A cut-off based approach for gene expression analysis of formalin- fi xed andparaf  fi n-embedded tissue samples Prashant K. Srivastava a,1 , Stefan Küffer a,1 , Benedikt Brors b,1 , Priyanka Shahi c , Li Li c ,Marc Kenzelmann a , Norbert Gretz c , Hermann-Josef Gröne a, ⁎ a Department of Cellular and Molecular Pathology, German Cancer Research Center, D-69120 Heidelberg, Germany b Department of Theoretical Bioinformatics, German Cancer Research Center, D-69120 Heidelberg, Germany c Medical Research Center, University of Heidelberg, D-68167 Mannheim, Germany a r t i c l e i n f o a b s t r a c t  Article history: Received 25 September 2007Accepted 4 March 2008Available online 19 May 2008 Microarray analysis of formalin- fi xed and paraf  fi n-embedded (FFPE) tissue seems to be of importance for thedetectionofmolecularmarkersetsinprostatecancer(PC).ThecompromisedRNAintegrityofFFPEtissueresultsin a high degree of variability at the probe level of microarray data as shown by degradation plot. We testedmethodsthatreducethevariabilitybyincludingallprobeswithin300nucleotides,within600nucleotides,oruptoacalculatedbreakpointwithreferencetothe3'-end.AcceptedPCpathwayssuchastheWntsignalingpathwaycould be observed to be signi fi cantly regulated within FFPE microarray datasets. The best representation of PCgene expression, as well as better comparability to meta-analysis and fresh-frozen microarray data, could beobtained with a 600-nucleotide cutoff. Beyond the speci fi c impact for PC microarray data analysis we propose acutoff of 600 nucleotides for samples for which the integrity of the RNA cannot be guaranteed.© 2008 Elsevier Inc. All rights reserved. Keywords: Formalin- fi xed paraf  fi n embedded tissueMicroarrayDegradation plotConstant cut-off Breakpoint Prostate cancer (PC) is the second leading cause of cancer-relateddeaths in males [1]. Prostate tumsrcenesis is a multistep process thatincludes the activation of oncogenes and the loss of tumor suppressorgenes that regulate cell proliferation and apoptosis [2]. PC has a longlatency; precursorlesionsthat representintermediate stages betweennormal and malignant cells can arise decades before the actualdiagnosisofcancer[3].Theprostate-speci fi cantigen(PSA)iscurrentlythebestserumindicatorforPC,withasensitivityandspeci fi cityof72.1and 93.2%, respectively [4]. However, the lack of additional markersimpedesearlydiagnosis,choiceoftreatment,andthepredictionoftheclinical outcome.High-throughput techniques such as microarray analysis of geneexpressioninfresh-frozen(FF)tissueshaveledtothediscoveryofasetof molecular markers associated with PC, such as  HOXB13  [5],  PCGEM1 [6,7],  TP63, AMACR  [8,9],  OR51E2  [10,11], and  TMPRSS2  [12]. However,FF tissue samples do not always provide suf  fi cient morphologicaldetails for a differential diagnosis of precancerous lesions, like severeprostatic intraepithelial neoplasia, and PC, for which array studiesmight be performed. Here a reliable diagnosis can be made only withformalin- fi xed and paraf  fi n-embedded (FFPE) tissues. In addition, insurgical pathology most samples are FFPE.The  fi xation process causes cross-linkage of nucleotides (nt) withproteinsandresultsincovalentmodi fi cationbytheadditionofmono-methylol groups to the bases of RNA. In addition to chemical modi- fi cations, RNA samples undergo severe mechanical stresses duringparaf  fi nembedding,whichaffecttheaveragelengthoftranscripts[13].Although several modi fi cations in the RNA extraction protocol havebeenmadeto improveRNAquality, thereliabilityof microarrayanaly-sis from FFPE samples continues to be a challenge [14 – 19]. We haveused PC as an example to establish a general protocol for microarrayanalysis of FFPE tissue samples.To optimize gene pro fi ling we used the Paradise reagent system(Arcturus, Mountain View, CA, USA) in combinationwith a specializedmicroarray chip that was speci fi cally developed to overcome RNAalterations in FFPE tissue.We performed a comparative analysis of FFPE tissue microarrays,with a comparable FF dataset using meta-analysis data, of four pub-licly available datasets as a reference of PC gene expression [20 – 23].A method for microarray analysis was explored that includes onlyprobes within the  fi rst 300 or 600 nt or up to a calculated breakpointof a transcript. Using this approach we were able to reduce the signalvariability on probe level that is caused by the shortened averagelength of RNA from FFPE samples.In comparison with meta-analysis data we have now shown thatan increase in sensitivity can be observed with the proposed methodin comparison to conventional approaches without  “ cutoff  ”  withoutlosing speci fi city and thereby increasing the reliability of FFPEmicroarray data analysis. Genomics 91 (2008) 522 – 529 ⁎  Corresponding author. Fax: +49 6221 424352. E-mail address: (H.-J. Gröne). 1 These authors contributed equally to this work.0888-7543/$  –  see front matter © 2008 Elsevier Inc. All rights reserved.doi:10.1016/j.ygeno.2008.03.003 Contents lists available at ScienceDirect Genomics  journal homepage:  Results Standard reference genes (SRG) Meta-analysis resulted in 1534 SRG. PC-speci fi c genes such as  PSA, AMACR, TP63, OR51E2,  and  HOXB13  showed to be signi fi cantlyregulated in SRG. All the studies that were taken into considerationfor meta-analysis performed microarray analysis on an older versionof Affymetrix chips (HGU95 series). Probes for the marker  PCGEM1 were not present on the HGU95 series microarray chip.Pathway analysis for SRG resulted in 30 signi fi cantly alteredpathwaystowhichwerefertoassigni fi cantreferencepathways(SRP). Degradation plot  A degradation plot shows average expression values for probeslocatedinintervalsof300ntstartingfromthe3'-endofatargetmRNAsequence. The degradation plot for FFPE samples can be divided intotwo parts: a steep decline between intervals 300 and 1200 and agradual decrease beyond interval 1200 in scaled signal intensity. The fi rst two intervals (300 and 600) have a difference of 0.01 and wereconsidered suitable distances for the implementation of constantcutoffs (Fig. 3). Signi  fi cance analysis of FFPE and FF datasets Signi fi cance analysis of microarray (SAM) and mixed-model analysis(MMA;SAS)wereemployedforperformingsigni fi canceanalysisovertheds_all,ds_300,ds_600,ds_break,andds_ffdatasets.Theanalysisresultedin 5871, 6608, 6617, 5530, and 4380 signi fi cantly regulated genes,respectively, for SAM and 2429, 1934, 2199, 309, and 4730 signi fi cantlyregulated genes, respectively, for SAS statistical methods. Analysis of thedatasets based on a cutoff (ds_300, ds_600, and ds_break) resulted in anincreased number of genes overlapping with the SRG (Table 1). Since theabsolute number of overlapping genes might be misleading, weintroduced the concept of true positive (TP), false positive (FP), truenegative(TN),andfalsenegative(FN).ForSAMamaximumsensitivityandspeci fi city can be observed with the ds_600 dataset (Table 1). Table 2 summarizes PC-speci fi c genes and their signi fi cance status withindatasets. SAM for the ds_600 dataset resulted in the maximum numberof PC-speci fi c genes. In contrast a minimum number of PC-speci fi c geneswere observed in the SAS ds_break dataset. Pathway analysis for signi  fi cantly regulated genes SRG were derived using SAM statistics. In a previous study [39], itwas shown that the comparability of methods (SAM and SAS)increases at the pathway level. Analyses were performed for theidenti fi cation of signi fi cantly altered pathways by implementingFisher's exact test for FFPE and FF. Genes found to be signi fi cantlyregulated for ds_all, ds_300, ds_600, ds_break, and ds_ff with SAMstatistics resulted in 128,126,149,128, and 58 and with SAS statisticsin 32, 36, 42, 2, and 59 signi fi cantly altered pathways, respectively.Table S2 lists signi fi cantly altered pathways with SRG and theirconcordance with the FFPE and the ds_ff datasets. We observed amaximum congruence of 19 of 30 SRP pathways with SAM for theds_600 dataset. The congruence between the two statistical methods,SAS and SAM, was 44% for ds_all and 70% for ds_ff, whereas for thecutoff datasets it was only between 0 and 26%. Table 3 summarizessensitivity and speci fi city at different signi fi cance levels for pathways Fig.1.  Exemplary Agilent Bioanalyzer pro fi les of RNA from four out of nine FFPE samples and its corresponding β -Actin 3 ’ /5 ’ -ratios. Pro fi les of (A) two individual non-cancerous and(B) twoindividualcanceroussamples areshown.(C) The shorteraveragelengthof a twostepampli fi ed cRNAfromaFFPE sample (left)is set in contrasttoaone-stepampli fi edcRNAfrom a FF sample (right).523 P.K. Srivastava et al. / Genomics 91 (2008) 522 – 529  signi fi cantlyalteredwithSAMforFFPEandFFdatasetswithrespect toSRP. Receiver operator characteristic curve (ROC) analysis at gene and pathway levels We performed ROC analysis for signi fi cantly regulated genes andpathways so as to identify the dataset on FFPE samples with anoptimal trade-off between sensitivity and speci fi city.Based on the de fi nitions of TP, FP, TN, and FN for pathway analysiswe performed a ROC analysis for all the FFPE datasets (Fig. 4). For theinitial points on the plot, the other datasets have comparable sensi-tivity and speci fi city but as we move along the x-axis we can observethat ds_600 excels among all the other datasets between points 0.1(90% speci fi city) and 0.35 (65% speci fi city). However, all datasetsfollow a similar trend beyond point 0.35 on the x-axis.ROCanalysiswasalsoperformedatthegenelevel(Fig.S1).Here,unlikeatthepathwaylevel,theds_ffdataset(areaunderthecurve(AUC)of78%)excels over all the other datasets, which range from 60 to 66%.  Analysis of variance (ANOVA) A 2×2 factorial ANOVA was performed to separate the effects due tothe methods (FFPE and FF) and the effects due to the disease (cancer ornormal). Method-speci fi c variations seem to be very high for all FFPEdatasets. The implementation of cutoffs (ds_300, ds_600, and ds_break)resulted in a number of probes that show signi fi cant method-speci fi ceffects. The datasets ds_break and ds_all had minimum and maximum Fig.2. WorkFlow.ProbesfromAffymetrixchips (HGU133A, U133_X3P,HGU95A andHGU95AV2) were annotated bysetting up standaloneBLASTagainst RNAdatabase (downloadedfrom NCBI). Based on the relative position with respect to the 3'-end of transcript/gene datasets were divided into 4 groups: ds_all, ds_300, ds_600 and ds_break respectively. SAMand SAS analysis was performed for the identi fi cation of signi fi cantly regulated genes, while Fisher's exact test was performed for pathway signi fi cance analysis. Meta-analysis wasperformed forthe fourpubliclyavailable datasets (seetextfordetail). The meta-analysisresults(SRGandSRP) wereused asreference setfor thecalculations ofsensitivity,speci fi cityand overlap calculations.524  P.K. Srivastava et al. / Genomics 91 (2008) 522 – 529  method-dependent variations, respectively. The result is summarized inTable S3. Discussion The gradual 5'-shortening and degradation of RNA from FFPEsamples is a known phenomenon. The low amount of RNA that isobtainedfromtissuesectionshastobeampli fi edtogeneratesuf  fi cientlabeled cRNA for microarray hybridization. A T7-based ampli fi cationusing random primers for the second cDNA synthesis is prone toshorten RNA fragments even more. To improve the issue of RNAdegradation a specialized microarray chip (Affymetrix, X3P series)thatcontainsadditionalprobessituatedwithin300ntfromthe3'-endof the transcripts can be used to optimize microarray results.OnanRNAchromatogramofanAgilentbioanalyzerthenarrowshiftedpeak of ampli fi ed cRNA from FFPE tissue compared to a FF sampleindicates a shorter average transcript length (Fig. 1C). The gradualdegradation of RNA results in a high variation at the probe level that canbe visualized in a degradation plot by plotting individual signals on theprobe level. This plot shows a steep decline between intervals 300 and1200 nt in terms of intensity values from 3 ′  to 5 ′  (Fig. 3). Due to thesevariationsnotallprobesrepresentingatranscript/genecanbeconsideredfor FFPE tissue microarray data analysis. Here, we have considered twoinitial intervals (300 and 600) since the difference in average intensitiesbetween these two points is very small.To reduce the variation within the transcript we implementedvarious cutoffs (ds_300, ds_600, and ds_break). A cutoff based on aconstant distancefromthe 3'-end (ds_300, ds_600) does not take intoconsideration that the extent of degradation varies across samples.Therefore a breakpoint analysis (ds_break) considering the variationfor each transcript across the samples with respect to RNA integritymight theoretically be better. ANOVA showed a clear decrease in thenumber of genes having method-speci fi c variation, which indicatesthat the implementation of probe  fi lters makes microarray resultsmore reliable (Table S3).SAMandSASanalyseswereperformedinparalleloverallFFPEandFFdatasets,fortheidenti fi cationofsigni fi cantlyregulatedgenes(Fig.2).Thenumber of genes identi fi ed to be signi fi cantly regulated varied betweenthe two methods employed. The differences in the numbers of signi fi cantly regulated genes obtained by the different methods wereaddressed by Wol fi nger and co-workers, who suggested that thesediscrepanciesmight bedue tothe different summarymethodsused [34].Comparing the signi fi cantly regulated genes of the differentdatasets to SRG, ds_ff had a very high number of genes overlappingwith SAMas wellas withSASanalysis(Table 1). ForFFPE,the constantcutoff datasets ds_300 and ds_600 showed an increase in number of overlapping genes without a loss of speci fi city with the implementa-tion of SAM statistics. The ds_600 dataset even had a maximumnumber of overlapping genes with SAM analysis. In contrast, with theds_break dataset, a decrease in sensitivityand no change in speci fi citywere observed. Implementing SAS statistics over the cutoff datasetsresulted in a signi fi cant decrease in the number of overlapping genes.In terms of sensitivity and speci fi city, a marginal decrease of 2 and 4%was observed for ds_600 and ds_300, respectively. One of thepotential reasons for the decrease in sensitivity with SAS could bethat the number of probes per transcript/gene was not suf  fi cient forperforming MMA. The larger decline (20%) in sensitivity for ds_breakcan be explained by the number of probes per transcript/gene thatvary across the samples. A second reason could be that the SRG wasobtained using SAM statistics. Hence, it has better concordance withthe genes obtained with signi fi cance analysis using SAM.ROC analysis was also performed at the gene level for the FF andFFPE datasets. The ds_ff dataset excels over all other datasets byhaving an AUC of 78% compared to FFPE datasets, which rangesbetween 60% for ds_300 and 65.8% for ds_all. On the x-axis betweenpoints 0 and 3.8, all FFPE datasets have a similar trend for sensitivityand speci fi city trade-off (Fig. S1).In the results from the ds_600 dataset all PC-speci fi c genes such as PSA,PCGEM1,TP63,OR51E2,TMPRSS2,AMACR, and HOXB13 werefoundto be signi fi cantly regulated (Table 2).Analysis using SAM and SAS differed in terms of numberof pathwaysshown to be signi fi cantly affected. The differences between the twomethodsatthepathwaylevelmightbeduetothefactthatthenumberof signi fi cantly altered pathways is directly proportional to the number of signi fi cantlyregulatedgenes.However,mostofthesigni fi cantlyregulatedpathways obtained with SAS were common to SAM. Compared to SRP,ds_all and ds_ff show better congruence between the methods than thedatasets ds_300, ds_600, and ds_break. A previous study [39] has  Table 1 Overlap for signi fi cantly regulated genes obtained with different datasetsds_all ds_300 ds_600 ds_break ds_ff Overlapping signi fi cantlyregulated genes SAM852 881 935 787 923Overlapping signi fi cantlyregulated genes SAS340 279 304 34 973Sensitivity SAM (%) 56 57 61 51 60Sensitivity SAS (%) 22 18 20 2 63Speci fi city SAM (%) 68 68 68 68 62Speci fi city SAS (%) 62 57 60 95 35Ifwe compare the FFPE datasets with SRG onecan observea clear increase inthenumberofgenesoverlappingwiththeimplementationofconstantcutoffs.Thehighestoverlapcanbeobservedbetweenconstantcutoffs(ds_300andds_600),whilethesmallestnumberof overlaps can be observed between the ds_break and the ds_all datasets. Fig.3. ThedegradationplotshowsaconstantdecreaseinnormalizedandscaledsignalintensityforallFFPEcancerandnon-cancermicroarraydataasonemovesalongtheX-axis(nt)orfurtherawayfrom3 ’ -end.The fi rsttwopointswerechosenforthecut-off  fi lters(300cut-offand600cut-off).Here,C1,C2,C3,C4andC5representcancersampleswhileN1,N2,N3and N4 represent samples from normal tissues.525 P.K. Srivastava et al. / Genomics 91 (2008) 522 – 529  demonstrated that comparability of different studies for PC microarraygene expression analysis as well as signi fi cance analysis by differentstatistics (SAM and MMA) improves at the pathway level. Table S4containsthelistofpathwaysthatwerealmost100%congruentwithSAMandSASanalysesandtheirsigni fi cancestatusforFFandFFPEdatasets.Forthissetofpathwaysaminimalmethod-dependenteffectcanbeobservedexceptfortheds_breakdataset.Duetothevariablenumberofprobespertranscript/gene across the samples SAS analysis could not be implemen-ted. It can be concluded that the method-speci fi c effect observed for SRPessentially exists. To compare all FFPE datasets, we considered SAM only.At the pathway level, an increase in sensitivity, as well asspeci fi city, can be observed for the ds_300 and ds_break datasetsfrom the signi fi cance level (  p -value) of 0.01 to 0.03 with respect tods_all. Beyond the signi fi cance level of 0.03, one can observe a slightdecline in the sensitivity and speci fi city percentages for ds_300 andds_break datasets. It is with the ds_600 dataset that a maximumincrease in sensitivity can be observed at all signi fi cance levels.Compared with SRP a maximum overlap of 19 pathways can beobtained with SAM on the ds_600 dataset. The other datasets (ds_all,ds_300,andds_break)resultedin16,16,and15overlappingpathways,respectively.Toreducethefalsepositiverate,ourinterestwastoobtainan optimal trade-off for sensitivity and speci fi city and not themaximum number of overlapping pathways. For FFPE datasets, ROCanalysis at the pathway and gene levels did not reveal a signi fi cantdifference in AUC (Figs. 4 and S1); although a slight increase in AUC atthe pathway level indicates that the dataset ds_600 has a slight edgeover all the other FFPE datasets.PathwayssuchasWnt,androgensignaling[40],p53signaling[41],and adenine monophosphate synthesis [42], which have all been shown toplayaroleinPCdevelopmentandprogression,aresigni fi cantwithallFFPEdatasets. Aberrant Wnt signaling has been shown to be important inprostatictumsrcenesisandthelevelofexpressionofWntpathwaygenesmightbeprognosticallyvaluable[43].Thebestrepresentationofpathwayswas obtained with the ds_600 and ds_ff datasets. Fig. 5 shows thesigni fi cantly up-regulated genes from the ds_600 and ds_ff datasetswithin the Wnt pathway, including  Frizzled, GSK3, Groucho, HDAC1, GBP, and NLK  incommonbetweenthetwodatasets.AdditionalWnt-activatedgenes such as CyclinD1  and PPAR δ were obtained with ds_600.In summary, the ability to use FFPE tissue samples for comparablemicroarray analysis would be a major advancement for retrospectivestudies of samples with a clinical follow-up. The improvement of isolation protocols and the approach for specialized microarray plat-formsenableresearcherstogenerateexpressionpro fi lesbasedonsmallamounts of RNA from FFPE samples with a relatively poor 3 ′  to 5 ′  ratio.Optimizing the analysis according to the integrity of RNA byapplying a  fi lter at probe level improves the ef  fi cacy for FFPEmicroarray data analysis.In theory ds_break would be the most optimal cutoff method tore fl ectRNAintegritystatusatprobelevel,althoughwedidnotobserveitas thebestmethodinourpresent study.Thisleadsusto believe thatthe statistical method behind breakpoint analysis certainly has thepotential for being improved with regard to microarray analysis.Several lines of evidence have now been provided that the bestrepresentation of PC as well as better comparability to standardreference genes and fresh-frozen microarray data can be obtainedwith the ds_600 dataset of FFPE. We therefore propose a constantcutoff of 600 nt to improve microarray data analysis for FFPE samplesfor which the integrity of the RNA cannot be guaranteed. Materials and methods Tissue and specimens FFPE and FF samples of human prostate adenocarcinoma were collected frompatients at the time of radical prostatectomy at the University Hospital Heidelberg andtheNephrologyCenter,HannoverschMünden.Afterradicalprostatectomy,independenttissues were either snap-frozen in liquid nitrogen or  fi xed in 4% buffered formalin (pH7.3) and embedded in paraf  fi nwithin 24 h. FFPE samples were collected between 2003and2005andstoredatroomtemperatureforatleast1yearbeforefurtherprocessing.FFtissueincludingnineprimarycancersandeightnoncanceroustissueswereexamined.Inaddition FFPE tissues of   fi ve primary cancers and four noncancerous areas wereanalyzed. Cancerous tissue was graded according to the Gleason scoring system by apathologist (H.-J.G.). Gleason scores of primary tumors ranged between 6 and 8. Tissue sectioning  Thirty to forty FF sections of 5  μ  m were cut at 20 °C under RNase-free conditions,transferred to an Eppendorf tube kept on ice, and snap frozen in liquid nitrogen. FFPEsections were cut at 5  μ  m and transferred to glass slides. Slides were dried at 37 °C for1handstoredat 80°Cuntilfurtheruse.The fi rst,themiddle,andthelastsectionsofFFand FFPE were stained by H&E to corroborate the diagnosis. RNA isolation, quality control, ampli  fi cation, and labeling of frozen tissue samples Tissue was taken up in guanidinium thiocyanate and homogenized by pipetting upand down. Further isolation was according to the protocol of Chomczynski and Sacchi[24]. RNA quality and quantity were analyzed with an RNA 6000 Nano LabChip on theBioanalyzer 2100 (Agilent Technologies, Palo Alto, CA, USA). Samples were labeled withthe One-Cycle Target Labeling Assay (Affymetrix, Santa Clara, CA, USA). Brie fl y, 2.5 μ  gof total RNA was used to generate double-stranded (ds) cDNA. The ds cDNA was puri fi edwith the Sample Cleanup Module (Affymetrix), and biotin-labeled cRNAwas generatedwith the GeneChip IVT Labeling Kit (Affymetrix). Ten micrograms of fragmented cRNAwas hybridized to the GeneChip Human Genome U133A Array (Affymetrix).Fragmentation, hybridization, washing, and staining were conducted according tothe manufacturer's recommendations. Labeled probes were denatured at 99 °C for  Table 3 Sensitivity and speci fi city at the pathway levelSensitivity Speci fi city Sensitivity Speci fi city Sensitivity Speci fi city Sensitivity Speci fi city Sensitivity Speci fi citySigni fi cance level (  p  value) 0.01 0.02 0.03 0.04 0.05ds_all 40 86.7 43.3 84.3 46.7 87 53.3 87.4 53.3 71.2ds_300 40 90 46.7 85.3 53.3 90.7 53.3 90.8 53.3 69.7ds_600 50 84 56.7 78.5 56.7 84.4 63.3 84.8 63.3 67.2ds_break 46.7 88.6 46.7 85.5 50 89 50 89 50 70ds_ff 20 92.4 26.7 92.4 26.7 95.4 26.7 95.4 26.7 86Sensitivity and speci fi city values at different signi fi cance levels across all datasets are shown. With the implementation of constant cutoffs we can observe a clear increase in thesensitivity at the cost of speci fi city. It is only with the ds_600 dataset that we observe an increase in sensitivity at different signi fi cance levels.  Table 2 PC-speci fi c genes and their signi fi cance and nonsigni fi cance across all datasetsGene ds_all ds_300 ds_600 ds_break ds_ff SRGSAM SAS SAM SAS SAM SAS SAM SAS SAM SAS SAM PSA  S S S S S S S NS NS NS S  AMACR  S S NA NA S S S S S S S PCGEM1  S NS S S S S S NS NS NS NA TP63  NS NS S NS S NS NS NS S S S OR51E2  S S NS NS S NS S NS S S S TMPRSS2  S S NS NS S NS S NS S NS NS HOXB13  S S NS NS S NS S NS S S SThe same number of markers can be observed with ds_all and ds_ff datasets. However,with the implementation of cutoff (ds_600 and ds_break) we are not losing any markergenes and yet we are improving the sensitivity for the ds_600 dataset (see text fordetails). S, signi fi cantly regulated with the respective analysis; NS, not signi fi cantlyregulated with the respective analysis; NA, not considered for the respective analysis.526  P.K. Srivastava et al. / Genomics 91 (2008) 522 – 529
Similar documents
View more...
Related Search
We Need Your Support
Thank you for visiting our website and your interest in our free products and services. We are nonprofit website to share and download documents. To the running of this website, we need your help to support us.

Thanks to everyone for your continued support.

No, Thanks