DCA: Dynamic Correlation Analysis

TianweiYu

Department of Biostatistics and Bioinformatics, Emory University, Atlanta, GA 30322,

USA. Email: tianwei.yu@emory.edu.

Abstract

In high-throughput data, dynamic correlation between genes, i.e. changing

correlation patterns under different biological conditions, can reveal important

regulatorymechanisms.Given thecomplexnatureofdynamiccorrelation,and the

underlying conditions for dynamic correlation may not manifest into clinical

observations, it is difficult to recover such signal from the data. Currentmethods

seekunderlyingconditionsfordynamiccorrelationbyusingcertainobservedgenes

as surrogates, which may not faithfully represent true latent conditions. In this

studywe develop a newmethod that directly identifies strong latent signals that

regulate the dynamic correlation of many pairs of genes, named DCA: Dynamic

Correlation Analysis. At the center of the method is a new metric for the

identification of gene pairs that are highly likely to be dynamically correlated,

withoutknowingtheunderlyingconditionsofthedynamiccorrelation.Wevalidate

theperformanceofthemethodwithextensivesimulations.Inrealdataanalysis,the

method reveals novel latent factors with clear biological meaning, bringing new

insightsintothedata.

Keywords:dynamiccorrelation,LiquidAssociation,latentvariables.

Introduction

The cellular system involves tens of thousands of genes/proteins that are tightly

regulated inacomplexnetwork(1-3). Interactionsandregulations inthenetwork

arehighlydynamic.Theychangesubstantiallyindifferentcelltypes,developmental

stages,orinresponsetoenvironmentalconditions(4).Geneexpressionandsimilar

typesofdata,suchasproteomicsandmetabolomicsdata,representoutcomesofthe

dynamic regulatory network. Changes in the underlying regulation patterns are

reflectedinthechangesingeneexpressionlevels,and/orchangesinthecorrelation

between genes. Many methods are available to analyze patterns in the gene

expressionlevels(5-8),while lessattentionhasbeenpaidtothestudyofdynamic

correlations.

Methods have been developed to find differential correlation patterns between

genesorgenesets,conditionedonagivenclinicalvariable(9-11).However,dynamic

correlationcanbemorecomplex.Underlyingcellularstatesmaynotmanifest into

clinicalobservations.Asthebiologicalsystemisregulatedinamodularmanner(12),

there could be multiple dynamic correlation conditions that govern different

functional groups of genes. Hence it is of interest to find unobserved dynamic

correlation conditions, which is a much harder problem. To this end, Li has

developed the Liquid Association (LA) approach, which uses a third gene as the

proxy of the dynamic correlation signal (13, 14). The method scans through all

possible gene triplets to find potential dynamic correlations. Similar approaches

thatutilizegenesaremediators(15,16),integrativeanalysisutilizingLA(17,18),as

wellassomestatisticaltheoryofLA(19)werelaterdeveloped.

Although focusing on gene-level dynamic correlations can reveal some important

localregulatorymechanisms,amoreglobalapproachtodynamiccorrelationcould

discovercriticalregulationmechanismsthatpenetratemultiplebiologicalprocesses,

orhelpidentifyhiddensub-groupsinthesamples.Tothisend,usingtheoriginalLA

orsimilarapproaches isnoteffectivedue to the followingreasons.First, scanning

through all possible triplets is computationally intensive. Second, a genome-scale

scan yields large numbers of LA gene triplets, causing difficulties in the

interpretation.Given theLAscore iscalculated inasymmetricmanneramongthe

threegenesinvolved,discerningwhichgenereflectscellularstatescouldbetricky.

Thirdandthemostimportant,thegenesthatserveassurrogatevariablesmaynot

begoodindicatorsoftrueunderlyingcellularstates.

In this study, our purpose is to find dominant dynamic correlation signals that

regulate the dynamic correlation of a large number of gene pairs. The biggest

difficulty is we do not know a priori which gene pairs have the relationship of

dynamiccorrelation.Wedesignanewmetric,namedLiquidAssociationCoefficient

(LAC), to effectively and efficiently screen all gene pairs for potential dynamic

correlations.Fromgenepairsthataremostlikelytobedynamicallycorrelated,we

provide a simple and straight-forward solution for quickly finding the latent

dynamic correlation signals. The procedure is named DCA: Dynamic Correlation

Analysis.WerefertothelatentsignalsfoundbyDCAasDynamicComponents(DCs).

Wedemonstratetheperformanceofthemethodusingextensivesimulations.Inreal

biologicaldatasets,wedemonstratethemethodcanidentifylatentsignalsthatare

biologicallymeaningful andnot foundby existingmethods. In amerged cell cycle

dataset, the method can find signals pertaining to the original experimental

grouping,aswellasbiologicalprocessesthatdifferentiatebetweentheexperiments.

IntheTCGAbreastcancer(BRCA)dataset,thenewmethodcanfindnewinteresting

subgroupsinthesubjectsthatarerelatedtopatientsurvivaloutcome.

Methods

Theoverallframework

Thedataisintheformofanexpressionmatrix,𝑮!×!,withpgenesintherowsandn

samples in the columns. Our assumption is that a portion of the gene pairs have

dynamiccorrelations,andtherearesomemajorlatentsignalsthatcanexplainmuch

of the variation in correlations among those gene pairs. Our purpose is to detect

suchdynamiccorrelationsignals.

Weassumethatallgenesarenormalizedtohavemean0andstandarddeviation1.

ThusthecovarianceandcorrelationbetweentwogenesXandYareequaltoE(XY).

First we assume we knowwhichm gene pairs have the relationship of dynamic

correlation. We address the selection of such gene pairs in the next sub-section.

Giventhesegenepairs,wecanconstructanewmatrix𝑩!×!,inwhichtheeachrow

is constructed bymultiplying the corresponding elements of a gene pairX andY,

𝑥!𝑦!, 𝑥!𝑦!,… , 𝑥!𝑦! .AgenecancontributetomultiplerowsoftheBmatrixifithas

dynamiccorrelationwithmultiplegenes.

For any z vector that is normally distributed, 𝑩𝒛 = 𝐿𝐴!, 𝐿𝐴!,… , 𝐿𝐴! ′ is

proportionaltotheLAscoreswithzbeingtheLAscoutinggeneoverall thepairs.

Fromaclusteringperspective,ifwefindclustersofrowsinthematrix𝑩,theneach

cluster shares a common LA scouting factor. Alternatively, from a principal

component perspective, 𝑩𝒛 ′ 𝑩𝒛 𝒛′𝒛is proportional to the sum of LA scores

squared over all the gene pairs. Finding a sequence of unit vectors𝒛that are

orthogonaltoeachotherandmaximizesthesumofLAscoressquaredrequiresthe

exactsamesolutionasconductingeigenvaluedecompositiononthematrix𝑩′𝑩.

Conceptually, other methods used to find latent factors, such as Independent

ComponentAnalysis (ICA) (20),SparsePrincipalComponentAnalysis (SPCA) (21),

ModularLatentStructureAnalysis(MLSA)(22),orvariousclusteringmethodscan

also be applied to the B matrix. In this manuscript we focus on the eigenvalue

decomposition approach. We note there is a caveat that this approach doesn’t

guaranteethatelementsof 𝒛willfollowthenormaldistribution.

Selectinginformativegenepairs

For the purpose of selecting informative gene pairs to find underlying dynamic

correlationsignals,wedefineameasurefordynamiccorrelationbetweenapairof

geneswithanunknowncondition factor, theLiquidAssociationCoefficient (LAC),

whichisthecorrelationcoefficientofthesquaredvaluesofthetwogenes,minusthe

correlationcoefficientoftheoriginalvaluessquared.

𝜁!,! = 𝑟 𝑔!!,𝑔!! − 𝑟! 𝑔! ,𝑔! ,

where𝑟()isthePearson’scorrelationcoefficient.Ithasbeenshownthatwhenboth

𝑔! and𝑔! follow the bivariate normal distribution with mean00 , and variance-

covariancematrix 1 𝜌!

𝜌! 1,thepopulationcorrelationcoefficientbetween𝑔!!and

𝑔!! isequalto𝜌!,whichmakestheabovequantityzero.

Alternatively, to reduce the impact of more extreme values, we can use the

correlation coefficient of the absolute values of the two genesminus the absolute

valueofthecorrelationcoefficient:

𝜁!,! = 𝑟 𝑔! , 𝑔! − 𝑟 𝑔! ,𝑔! .

WecomputethematrixofLACvaluesforallpairsofgenes.Noticethecomputational

cost is on the same scale as computing the pairwise correlationmatrix.We then

select the 𝑖, 𝑗 pairs whose LAC values are above a certain percentile of all the

valuesinthematrix.

After selecting the top 𝑖, 𝑗 pairs,weconstruct theBmatrix, inwhicheachrow is

constructedfromaselectedpairofgenes.Forexample,if𝑔! and𝑔! areselectedasa

pair of informative genes, then the corresponding row of the new matrix is

𝑔!!𝑔!!, 𝑔!!𝑔!!,… , 𝑔!"𝑔!" .Inthisstudy,weuseeigenvaluedecompositionofB’Bto

extract latent factors, and varimax rotation (23) to improve the interpretability of

thelatentfactors.

Selectinggenepairsassociatedwithalatentfactor

We first calculate theLAC coefficients for all pairs of genes, and select genepairs

withLAC coefficients belonging to a top percentile (20% in this study). We then

calculate their LA scores with the latent factor. Heuristically, we model the

distributionofLAscoresasamixture,withadominantsplit-normalcomponentin

thecenterrepresentinggenepairswithnorelationtothelatentfactor,i.e.thenull

distribution. We apply the local false discovery (fdr) approach to calculate the

posteriorprobabilitythatagenepairbelongstothenon-nulldistribution(24),and

threshold the fdrvalues toselectgenepairs thataredynamicallycorrelatedgiven

thelatentfactor.

Findingbiologicalprocessesassociatedwithalatentfactor

For functional interpretation,weusegeneontology (GO)biologicalprocesses.We

firstselectasetofrepresentativeGObiologicalprocesstermsthatareofreasonable

size and relatively small overlaps, following an existing procedure that considers

both theontology structure and thenumberof genes assigned to each term (25).

Fortheyeastdata,weselect172biologicalprocesseswith50~1000assignedgenes

each, covering 5334 genes in total. For the human data, we select 423 biological

processeswith100~1000assignedgeneseach,covering14414genesintotal.From

thegenepairsassociatedwitheachlatentfactor,weconducttwotypesofanalyses:

Within-process dynamic correlation. For each biological process, we count the

occurrenceofgenepairsinwhichbothgenesfallintotheprocess.Wealsocalculate

theexpectednumberofsuchgenepairsifallthegenepairswererandomlydrawn.

Wecalculatethefold-changebytakingtheratioofobservedcountv.s.theexpected

count,andp-valueusingthebinomialdistribution.

Between-process dynamic correlation. For each pair of selected biological

processes,wefirstremovetheiroverlappinggenes.Wethencounttheoccurrenceof

gene pairs in which the two genes fall into the two processes respectively, and

calculate the expected number of such gene pairs if all the geneswere randomly

drawn.Afterthresholdingthefoldchangeandp-valuetoselectpairsofprocesses,

wevisualizetheresultingnetworkusingCytoscape(26).

ResultsandDiscussion

IllustrationoftheLiquidAssociationCoefficient(LAC)

Inthisstudyanewmetricisdefinedtorankallpairsofvariablesinthedatamatrix.

ThepurposeoftheLACistohelpidentifygenepairsthataremostlikelytohavethe

relationshipofdynamiccorrelation,withoutknowingtheunderlyingconditionsof

the dynamic correlation. Gene pairs with such relations should receive high LAC

score,whileothergenepairs,eitherindependentorcorrelated,shouldreceivelow

scores.

The LAC requires all variables to have mean zero and standard deviation 1. As

illustrated in Figure 1, if both variables X and Y follow the standard normal

distributionmarginally, and one-third of the (X,Y) pairs are positively correlated,

one-third of the (X,Y) pairs are negatively correlated, and another one-third

uncorrelated, then the absolute values will be positively correlated, and the LAC

tends tobe large (Fig. 1, left column).On theotherhand,whenX andY are truly

independentorsimplycorrelated,theLACtendstobesmall.

Wefurtherconductalargersimulationstudytoexaminetheempiricaldistribution

of LAC under different circumstances. As illustrated in Figure 2, when the two

variablesaredynamicallycorrelated,thedistributionoftheLACscoreiscenteredat

apositivevalue(Fig.2,bluecurves).Thehigherthecorrelationlevel,thehigherthe

mean(Fig.2,lefttorightpanels).Thehigherthesamplesize,thelessthespread(Fig.

2,different linetypes).At thesametime, in the independentandcorrelatedcases,

theLACscoresarecenteredaroundzeroifthefirstdefinitionofLACisused.Using

theseconddefinition,theLACisstillcenteredaroundzerointheindependentcase,

andthecenterisnegativeinthecorrelatedcase(Fig.2,lowerpanels).

Weconductedanextensivesimulationstudytoevaluatethemethod’scapabilityto

recover latent dynamic correlation signals. Please refer to the Supporting

Information,section1fordetails(SupportingFigures1~3).Overall,themethodcan

recover the hidden dynamic correlation signal when the sample size and signal

strengthissufficient.

DCAextractssignalsthatdifferentiateexperimentsfromthemergedcellcycledata

Wefirstanalyzethewell-studiedSpellmancellcyclegeneexpressiondata(27).The

datasethasbeenanalyzedbymanyauthors.Thepurposeoftheanalysishereisto

demonstrate that DCA can extract information that is clearly meaningful, and

providesnovelbiologicalinsights.

Thecellcycledataset includes fourtime-seriesexperimentsof theyeastcellcycle,

eachusingadifferentmethodofsynchronization.Thetotaldimensionis6178genes

by 73 samples. Missing values were imputed by the K-nearest neighbor (KNN)

method(28).Whenallfourtimeseriesdatasetsarecombinedintoasingledataset,

traditionalmethodssuchasPCAandSPCA(21)extractsignals thatareconsistent

across the four time series (Supporting Figures 4 and 5), but not signals that

separatethefourtimeseries,exceptthefirstPCthatcapturesanoscillatingsignal

whichisanartifactintheCDC15timeseriesdata(29).

Applying DCA to the combined cell cycle data yields factors that are distinctly

different.Most of the Dynamic Components (DCs) clearly differentiate one of the

fourtimeseriesfromtherest(SupportingFigure6).Forafulllistoffactorplotsand

biologicalprocessesassociatedwitheach factor,please refer toSupportingFile2.

Herewefocusourdiscussiononthreeofthefactors.

The first DC has high scores for samples from the CDC15 experiment only. It has

been documented that an oscillating signal is present in the CDC15 data across

manygenes (SupportingFigure7), causinganelevated levelofcorrelationoverall

(29). ThefirstDCreflectsthissignal.Atthesametime,genepairsassociatedwith

thisDCarenotclearlyassociatedwithanybiologicalfunction,asreflectedinthefact

that no biological function pairswere found at the threshold of p=0.001 and fold

change=2.Thisisexpectedgiventhefactthattheoscillatingsignalisnotbiologically

meaningful.

ThesecondDConlyhasextremescores forsomeof thesamplesof theelutriation

experiment.AcloserexaminationrevealstheDCshowsasine-wavepatterninthe

elutriationsamples(Figure3).Anexaminationofthedatarevealsastrongdynamic

correlation pattern between genes associatedwith thisDC (Supporting Figure 8).

Selecting biological processes pairs that have excessive dynamic correlation links

between them, we find that the processes are focused on rRNA biogenesis and

ribosomeassembly.Muchmorepositive/negativecorrelationsareshownbetween

genesinthesebiologicalprocesseswhentheDC2scoreislow,whichcorrespondto

halfof the samples in theelutriationexperiment (SupportingFigure8). Whileall

the other three experiments are based on block-and-release cell cycle

synchronization,theelutriationprocessseparatessynchronizedcellsbasedontheir

size,shapeandmass(30).Theresultshereindicatethatproteinbiosynthesistend

to be better synchronized in the elutriation samples compared to the other three

experiments.

For the fifth DC, samples in the CDC28 experiment have lower scores, while the

alphafactorsampleshavehigherscores,withasmallermagnitude(Figure3).This

indicatesthatsomegenepairshaveareversecorrelationpatternbetweenthetwo

experiments,whichisintriguinggivenbothexperimentsusedblock-and-releaseto

synchronize cells. Some more recent studies have shed light on the metabolic

behavioroftheyeastcellsunderthealphafactororCDC28cellcyclearrest.Under

thealphafactortreatment,thecentralmetabolic fluxesareatahighlevel,andthe

cellularmetabolismtendtoberespiratoryevenwhenglucoseisabundant(31).The

cell cycle CDK Cdc28 regulates both the cell division processes and metabolic

processes.UndertheCDC28inhibition,thecellsaccumulateglycogenandtrehalose

toextremelyhighlevels(32).Giventhedifferentcharacteristicsofthetwocellcycle

arrestmechanisms,itisunderstandablethatafterthereleaseofcellcyclearrest,the

cellsproceedfromverydifferentmetabolicsituations,andmetabolismwilladaptto

thosesituations.SupportingFigure9showsgenesassociatedwithDC5,wherewe

canobserveaverystrongpattern in theCDC28samples, andaweakerpattern in

thealpha factorsamples.Functionally,weobserve thehighlyconnectedbiological

processes mostly involve small molecule metabolism and transport (Figure 4b).

TwotypicalpairsofgenesareshowninFigure4c,wherecleardynamiccorrelation

isobserved.

Overall, unlike traditional methods such as PCA and SPCA that identify

commonalities, the DCA approach tend to find signals that differentiate the four

underlying experiments, and reveals some important biological processes that

behavedifferentlybetween the experiments.Given the existing knowledgeon the

dataset, these results validate that DCA extract new andmeaningful information.

However, inmostotherapplications, informationsuchassamplegroupingarenot

available. We next examine the TCGA breast cancer (BRCA) dataset to see if the

methodcanextractanynewinsightsfromthedata.

DCAbringsnewinsightsintotheTCGABreastCancerdata

The data contains the measurement of 20532 genes by deep sequencing in 762

subjectswithbreastcancer.Afterremovinggeneswith>20%zeroreadings,17728

genes remain in the study. Similar to the yeast cell cycle data, the DCA captures

signalsthataredistinctfromtraditionalmethods.Herewefocusourdiscussionon

threeoftheDCs,astheyareclearlylinkedtoestrogenreceptor(ER)status(Figure

5a, Supporting Figure 10). DC1 largely separates ER-positive and ER-negative

samples,whichagreeswiththesecondprincipalcomponentverywell(Figure5b).

Ontheotherhand, inthespacespannedbyDC3andDC7,ER-positivesamplesare

tightly clustered in themiddle,while part of the ER-negative samples are spread

widely(Figure5a,SupportingFigure10).NoPCscaptureasimilarstructureinthe

data(SupportingFigure11).

Further analyses show that among the ER-negative subjects, those with more

extreme scores in eitherDC3orDC7 showadifferent survival characteristic than

thoseinthecenter(Figure5c).Thesubjectswithmoreextremescorestendtohave

a higher chance of dying earlier, while in long follow-ups the remaining subjects

tendtosurvivelonger,albeitsupportedbyrelativelyfewdatapoints.

Functionally, the biological processes that show excessive dynamic correlations

conditionedonDC3arecenteredaroundtwomainthemes(Figure6a).Thefirstis

protein sumoylation and stress response. Sumoylation is a post-translational

modificationthatoftenoccurs inresponse tocellularstress(33).Manyoncogenes

and tumor suppressors are functionally related to sumoylation (34). The second

mainthemeiscelldifferentiationandtissuedevelopmentthatarerelatedtoseveral

typesof tissues, indicatingadysregulation in thecells.GenesassociatedwithDC3

mainlyfallintotwogroupsthatexhibitinversecorrelationwhenDC3scoreislow,

andlowexpressionwhenDC3scoreishigh(SupportingFigure12).

The biological processes associated with DC7 are mostly immune response

processes (Figure 6b). Patterns of immune cell infiltration has been linked to the

prognosis and treatment response of breast cancer (35). An examination of the

genesassociatedwithDC7revealsthatoverhalfofsuchgenesarelowlyexpressed

when DC7 score is more extreme. A smaller portion of the genes are lowly

expressed when DC7 score is low, and highly expressed when DC7 score is high

(SupportingFigure13). In thissituation, themethod in factdetectsa latent factor

thathasconditionalmean-shifteffectsonthe immunegenes,whichwasdiscussed

byHoetal(19).Thechangedexpressionpatternsofmostlyimmune-relatedgenes

in these samples are likely reflective of a certain immune cell infiltration pattern

thathasimplicationsinprognosis.BesidethethreeDCsthatwediscusshere,most

oftheotherDCsshowclearfunctionalimplications,butrequireextrastudybeyond

this manuscript to elucidate their biological meaning. The full results are in

SupportingFile3.

Overall,asanewunsupervisedlearningmethodforhighdimensionaldata,DCAcan

extract new and useful information from the data. It complements existing

dimension reduction methods to reveal more internal structure in the data that

could lead to new biological discovery. The method is straight-forward, and the

computation is efficient. The R package is available at https://cran.r-

project.org/web/packages/DCA/index.html.

Acknowledgments

ThisworkwaspartiallysupportedbyNIHgrantsU19AI090023andU19AI057266.

TheauthorthankMr.YunchuanKong,Dr.JianKang,andDr.PeterSongforhelpful

discussions.

References

1. Barabási A-L (2007) Network medicine--from obesity to the "diseasome".TheNewEnglandjournalofmedicine357:404-407.

2. BarabásiA-L,GulbahceN,&LoscalzoJ(2011)Networkmedicine:anetwork-basedapproachtohumandisease.Naturereviews.Genetics12:56-68.

3. ChanSY&LoscalzoJ(2012)Theemergingparadigmofnetworkmedicineinthestudyofhumandisease.Circulationresearch111:359-374.

4. IdekerT&KroganNJ(2012)Differentialnetworkbiology.Molecularsystemsbiology8:565.

5. Rapaport F, et al. (2013) Comprehensive evaluation of differential geneexpressionanalysismethodsforRNA-seqdata.GenomeBiol14(9):R95.

6. Eren K, Deveci M, Kucuktunc O, & Catalyurek UV (2013) A comparativeanalysisofbiclusteringalgorithmsforgeneexpressiondata.BriefBioinform14(3):279-292.

7. Andreopoulos B, An A, Wang X, & Schroeder M (2009) A roadmap ofclustering algorithms: finding a match for a biomedical application. BriefBioinform10(3):297-314.

8. Meng C, et al. (2016) Dimension reduction techniques for the integrativeanalysisofmulti-omicsdata.BriefBioinform17(4):628-641.

9. Gill R, Datta S, & Datta S (2010) A statistical framework for differentialnetworkanalysisfrommicroarraydata.BMCbioinformatics11:95.

10. YuT&BaiY(2011)Capturingchangesingeneexpressiondynamicsbygenesetdifferentialcoordinationanalysis.Genomics98(6):469-477.

11. ChoSB,KimJ,&KimJH(2009)Identifyingset-wisedifferentialco-expressioningeneexpressionmicroarraydata.BMCBioinformatics10:109.

12. WagnerGP,PavlicevM,&CheverudJM(2007)Theroadtomodularity.NatRevGenet8(12):921-931.

13. LiKC (2002)Genome-widecoexpressiondynamics: theoryandapplication.Proceedings of the National Academy of Sciences of the United States ofAmerica99(26):16875-16880.

14. LiKC,LiuCT,SunW,YuanS,&YuT(2004)Asystemforenhancinggenome-wide coexpression dynamics study.Proceedingsof theNationalAcademyofSciencesoftheUnitedStatesofAmerica101(44):15561-15566.

15. Boscolo R, Liao JC, & Roychowdhury VP (2008) An information theoreticexploratorymethod for learning patterns of conditional gene coexpressionfrommicroarraydata.IEEE/ACMTransComputBiolBioinform5(1):15-24.

16. Chen J, Xie J, & Li H (2011) A penalized likelihood approach for bivariateconditional normal models for dynamic co-expression analysis. Biometrics67(1):299-308.

17. Yan Y, et al. (2017) Detecting subnetwork-level dynamic correlations.Bioinformatics33(2):256-265.

18. Wang L, et al. (2017) Meta-analytic framework for liquid association.Bioinformatics.

19. HoYY,ParmigianiG,LouisTA,&CopeLM(2011)Modelingliquidassociation.Biometrics67(1):133-141.

20. Hyvarinen A & Oja E (2000) Independent component analysis: algorithmsandapplications.NeuralNetw13(4-5):411-430.

21. ZouH,HastieT,&TibshiraniR(2006)Sparseprincipalcomponentanalysis.JournalofComputationalandGraphicalStatistics15(2):265-286.

22. YuT (2010)An exploratorydata analysismethod to revealmodular latentstructuresinhigh-throughputdata.BMCbioinformatics11:440.

23. Bernaards CA & Jennrich RI (2005) Gradient Projection Algorithms and Software forArbitrary Rotation Criteria in Factor Analysis. EducationalandPsychologicalMeasurement65:676-696.

24. EfronB(2004)Large-scalesimultaneoushypothesistesting:Thechoiceofanullhypothesis.JAmStatAssoc99(465):96-104.

25. YuT,SunW,YuanS,&LiKC(2005)Studyofcoordinativegeneexpressionatthebiologicalprocesslevel.Bioinformatics21(18):3651-3657.

26. Shannon P, etal. (2003) Cytoscape: a software environment for integratedmodelsofbiomolecularinteractionnetworks.GenomeRes13(11):2498-2504.

27. Spellman PT, et al. (1998) Comprehensive identification of cell cycle-regulated genes of the yeast Saccharomyces cerevisiae by microarrayhybridization.MolBiolCell9(12):3273-3297.

28. Troyanskaya O, et al. (2001) Missing value estimation methods for DNAmicroarrays.Bioinformatics17(6):520-525.

29. LiKC,YanM,&YuanSS(2002)Asimplestatisticalmodel fordepictingthecdc15-synchronized yeast cell-cycle regulated gene expression data. StatSinica12(1):141-158.

30. Smith J, Manukyan A, Hua H, Dungrawala H, & Schneider BL (2017)SynchronizationofYeast.MethodsMolBiol1524:215-242.

31. Williams T, Peng B, Vickers C, & Nielsen L (2016) The Saccharomycescerevisiaepheromone-responseisametabolicallyactivestationaryphaseforbio-production.MetabolicEngineeringCommunications3:142-152.

32. Zhao G, Chen Y, Carey L, & Futcher B (2016) Cyclin-Dependent Kinase Co-OrdinatesCarbohydrateMetabolismandCellCycle inS. cerevisiae.MolCell62(4):546-557.

33. WilkinsonKA&HenleyJM(2010)Mechanisms,regulationandconsequencesofproteinSUMOylation.BiochemJ428(2):133-145.

34. Eifler K & Vertegaal AC (2015) SUMOylation-Mediated Regulation of CellCycleProgressionandCancer.TrendsBiochemSci40(12):779-793.

35. Ali HR, Chlon L, Pharoah PD,Markowetz F, & Caldas C (2016) Patterns ofImmuneInfiltrationinBreastCancerandTheirClinicalImplications:AGene-Expression-BasedRetrospectiveStudy.PLoSMed13(12):e1002194.

Figures

Figure 1. Illustration of liquid association coefficient (LAC). Left column:dynamiccorrelationwithanunknownconditioningfactor.Whenthefactorislow,xand y are negatively correlated; when the factor is high, x and y are positivelycorrelated. Second left column: independent case. Right two columns: correlatedcase.Inallthecases,themarginaldistributionofXandYarestandardnormal.

Figure 2. Empirical distributions of LAC score under conditions of dynamiccorrelation, simple correlation, or independence. The densities are based on1000 simulations. In the dynamic correlation cases, one-third of the data pointsfollow a bivariate normal distribution with mean 00 and variance-covariance

matrix 1 𝜌𝜌 1 ,one-thirdfollowabivariatenormaldistributionwithmean 00 and

variance-covariancematrix 1 −𝜌−𝜌 1 , andanotherone-third follow independent

standard normal distributions. In the correlated case, all data points follow abivariate normal distribution with mean 00 and variance-covariance matrix1 𝜌𝜌 1 .

Figure3.SomeexampleDynamicComponentsfromthecellcycledata.Colors:thefourcellcycleexperiments.Red:alphafactor;green:CDC15;blue:CDC28;purple:elutriation.

0 10 20 30 40 50 60 70

-0.1

0.0

0.1

0.2

0.3

DC 1

Index

DC

sco

re

0 10 20 30 40 50 60 70

-0.4

-0.2

0.00.1

DC 2

Index

DC

sco

re

0 10 20 30 40 50 60 70

-0.4

-0.2

0.0

0.2

DC 5

Index

DC

sco

re

60 65 70

-0.4

-0.2

0.00.1

Index

DC

sco

re

Figure4.Biologicalprocesspairswithexcessivedynamiccorrelationsrelatedto DCs 2 and 5. Gene pairswere selected using fdr threshold of 0.01. Biologicalprocesspairswereselectedusingap-valuethresholdof0.001andfold-changeof2.For simplicity, only nodeswith connections above a certain threshold are shown.Node sizes reflect the total number of connections of each node. (a) Biologicalprocesspairsassociatedwiththe2ndDC.(b)Biologicalprocesspairsassociatedwiththe5thDC. (c)ExampleplotsofgenepairswithLArelationwithDC5.Redpoints:samples inthe lower33%ofDC5score;bluepoints:samples intheupper33%ofDC5score.

(a)(b)

(c)

Figure5.ResultsfromtheTCGABRCAdataset.(a)ScatterplotsofDC1,DC3,andDC7 scores. The points are colored based on the ER status of the subjects. DC1separates ER+ and ER-, while DC3 and DC7 have awide spread only for the ER-subjects. (b)DC1capturessimilar informationas thesecondprincipalcomponent.(c)SurvivalcurvesoftheER-negativesubjects,red:absolutefactorscore>0.05.

(a)

(b)

Figure6.Biologicalprocesspairswithexcessivedynamiccorrelationsrelatedto DCs 3 and 7. Gene pairswere selected using fdr threshold of 0.01. Biologicalprocesspairswereselectedusingap-valuethresholdof0.001andfold-changeof3.For simplicity, only nodeswith connections above a certain threshold are shown.Node sizes reflect the total number of connections of each node. (a) Biologicalprocesspairsassociatedwiththe3rdDC.(b)Biologicalprocesspairsassociatedwiththe7thDC.