Grouping of histone deacetylase inhibitors and other toxicants disturbing neural crest migration by transcriptional profiling

Grouping of histone deacetylase inhibitors and other toxicants disturbing neural crest migration by transcriptional profiling

Nadine Dreser

^a,1

, Bastian Zimmer

^b,c,1,

*, Christian Dietz

^d

, Elena Su¨gis

^e

, Giorgia Pallocca

^a

, Johanna Nyffeler

^a

, Johannes Meisig

^f,g

, Nils Blu¨thgen

^f,g

, Michael R. Berthold

^d

,

Tanja Waldmann

^a,2

, Marcel Leist

^a,2

aDoerenkamp-ZbindenChairofInVitroToxicologyandBiomedicine,UniversityofKonstanz,Konstanz,Germany

bCenterforStemCellBiology,Sloan-KetteringInstitute,NewYorkCity,NY,USA

cDevelopmentalBiologyProgram,Sloan-KetteringInstitute,NewYorkCity,NY,USA

dLehrstuhlfu¨rBioinformatikundInformationMining,UniversityofKonstanz,Konstanz,Germany

eInstituteofComputerScience,UniversityofTartu,Tartu,Estonia

fInstituteofPathology,Charite´-Universita¨tsmedizin,10117Berlin,Germany

gIntegrativeResearchInstitutefortheLifeSciencesandInstituteforTheoreticalBiology,HumboldtUniversita¨t,10115Berlin,Germany

1. Introduction

Thecapacityofneuralcrestcells(NCC)tomigratecorrectlyto different parts of the fetus where theydifferentiate to various tissues(e.g.bone,cartilage,neurons,melanocytes,etc.)isessential

for normal human development (Dupin and Sommer, 2012).

Defects in neuralcrestfunction and differentiationcan lead to severebirthdefectsanddiseasessuchascleftpalate,Hirschsprung diseaseorCHARGEsyndrome(KeyteandHutson,2012).Several chemicals are known to disturb the NCC development and migration and thereby cause developmental defects (Di Renzo et al., 2007; Fuller et al., 2002; Menegola et al., 2000).

Acomprehensivehazardassessmentapproachrequirestherefore testsystemsthatdetectdisturbancesofNCCfunction.

Tomonitorwhetherchemicalsdisturbthemigrationcapacityof NCC, theMINC (migration of neuralcrest cell) assay has been developed.Thistestsystemassesses,howmanycellsre migrate Keywords:

Neuralcrest Migration Read-across Chemicalgrouping Geneexpression HDACinhibitors Pesticides Metals PCBs

Pathwaysoftoxicity Toxicology

ABSTRACT

Functionalassays,suchasthe‘‘migrationinhibitionofneuralcrestcells’’(MINC)developmentaltoxicity test,canidentifytoxicantswithoutrequiringknowledgeontheirmodeofaction(MoA).Here,wewere interested,whether(i)inhibitionofmigrationbystructurallydiversetoxicantsresultedinaunified signature oftranscriptional changes; (ii) whetherstatistically identifiedtranscript patterns would informoncompoundgroupingeventhoughindividualgeneswerelittleregulated,and(iii)whether analysisofasmallgroupofbiologically relevanttranscriptswouldallowthegroupingofcompounds according to their MoA.We analyzed transcriptsof35 ‘migrationgenes’ aftertreatment with16 migration inhibitingtoxicants.Clustering,principalcomponentanalysisandcorrelationanalysesofthe datashowedthatmechanisticallyrelatedcompounds(e.g.histonedeacetylaseinhibitors(HDACi),PCBs) triggeredsimilartranscriptionalchanges,butgroupsofstructurallydiversetoxicantslargelydifferedin theirtranscriptional effects.Lineardiscriminantanalysis(LDA) confirmedthespecificclusteringof HDACiacrossmultipleseparateexperiments.SimilarityofthesignaturesoftheHDACitrichostatinAand suberoylanilidehydroxamicacidtotheoneofvalproicacid(VPA),suggestedthatthelattercompound actsasHDACiwhenimpairingneuralcrestmigration.Inconclusion,thedatasuggestthat(i)agiven functionaleffect (e.g.inhibitionofmigration) canbe associated with highlydiversesignatures of transcriptchanges;(ii)statisticallysignificantgroupingofmechanistically relatedcompoundscanbe achievedonthebasisoffewgeneswithsmallregulations.Thus,incorporationofmechanisticmarkersin functional in vitro tests may support read across procedures, also for structurally un related compounds.

* Correspondingauthorat:SloanKetteringInstitute,1275YorkAve,RRL945, NewYork,NY10065,USA.

E-mailaddresses:bastian.zimmer@uni-konstanz.de, zimmerb@mskcc.org(B.Zimmer).

1Theseauthorssharefirstauthorship.

2Theseauthorsshareseniorauthorship.

Konstanzer Online-Publikations-System (KOPS) URL: http://nbn-resolving.de/urn:nbn:de:bsz:352-310991 Erschienen in: NeuroToxicology ; 50 (2015). - S. 56-70

https://dx.doi.org/10.1016/j.neuro.2015.07.008

intoacell freearea(scratch)withinaneuralcrestcellmonolayer.

Severaltoxicantsknowntoaffectneuralcrestcellsinvivohave been tested positive in this test. They include triadimefon, mercury,lead andtheantiepileptic drugvalproicacid(Zimmer et al., 2012). Using the inhibition of migration capacity as an endpoint for toxicity testing seems promising for detection of chemicals potentially hazardous to human early development.

Althoughsomebiologicalpathways(e.g.actinpolymerizationor src signaling) have been identified to be required for NCC migration,it isnotclear,whichofthemisaffectedbytoxicants orwhetheralltoxicantsthataffectNCCmigration,convergeonthe same biological pathway. Thus the underlying mode of action (MoA) and the associated adverse outcome pathways (AOP) remainunknownand theassay,asmostofthecurrent invitro andinvivotoxicityassays,isstillablackboxapproach.

First attempts havebeen undertakento combine traditional assays that are based on so called apical endpoints with measurementsofintermediateevents,suchaschangesofmRNA, proteins or metabolites. For instance, it has been shown that morphological deficits triggered by triazoles in the rat whole embryoculturewereaccompaniedbycharacteristictranscriptome signatures (Robinson et al., 2012), and neuronal cell death triggered by the toxicant 1 methyl 4 phenylpyridinun was precededbypronouncedchangesofmetabolitesand transcripts (Krugetal.,2014).

Such approaches mightallowthedetectionof keyevents of AOPs,definitionofbiomarkers(Kuegleretal.,2010)ordelineation of an entire MoA/AOP (Bal Price et al., 2015b). On this basis relevantassays couldbecombinedintotoa moreefficienttest battery(Piersma,2014;Piersmaetal.,2013;Zimmeretal.,2014).

One unsolved issue in this process is how a final adverse outcome (e.g. disturbed migration in the MINC) links to intermediate changes inside the cells. One assumption is that toxicantscausingthesametypeanddegreeofcomplexadverse outcome, also lead to similar changes of relevant transcripts, proteinsorotherpotentialbiomarkersoftoxicity(Blaauboeretal., 2012).

Anoppositehypothesiswouldbethateachtoxicanttriggersits ownparticularsetofpathwayalterations,andthatmanydifferent suchchangesofcellfunctioneventuallyleadtothesameadverse outcome.Signaturesofbiologicaldisturbanceswouldinsuchcases be different for each toxicant. In between these extremes, a realisticsituationcouldbethatalimitednumberofkeyeventsof differentbiologicalpathwaydisturbancesleadtothesameadverse outcome.

Thus,toxicantssharingamodeofactionwouldtriggersimilar changesofintermediate cellularmarkers.Severaltoxicitypath waysmaylead tothe sameadverse outcome (Patlewiczet al., 2014),sothatonefinaloutcomecouldbelinkedtodifferent(more thanone,butmostlikelynothundreds)signaturesofintracellular changes. Thiswould allowgrouping of compoundsalong their specificdisturbancepatternsorbiomarkers.Suchgroupsofupto 20 compounds have for instance been observed for the 148 compoundstestedintheTGGATESproject(Grinbergetal.,2014).

Groupingalongmechanisticmarkershasalsobeendoneforthe ToxCast phase I chemicals (Kleinstreuer et al., 2014). Such information could guide biologically driven read across proce dures in hazard assessment (Bal Price et al., 2015a) to group compoundsbasedonsimilaritiesofMoA.Thiswouldcomplement traditionalstructure basedapproaches.

Inthisstudywefirstidentifieddiversegroupsoftoxicantsthat disturbthemigrationofneuralcrestcellsintheMINCassay.Using asetof35migration relatedgenes,wemonitored,byRT qPCR, transcriptional changes triggered by these toxicants. Notably, highlysignificantgroupingandcorrelationoftranscriptsignatures was identified on the basis of small transcript changes, not

significant on the single gene level. Grouping of toxicants accordingtotheirgeneexpressionprofilesidentifiedthegroup ofHDACinhibitorsashithertounrecognizeddrugclassaffecting NCCmigration.Thestudyshows,thatthetestsystemgiveshighly reproducible results across different biological replicates, not onlyfortheprimaryapicalendpoint(i.e.inhibitionofmigration capacity), butalso fortranscriptchanges, as mechanisticend points.Ingeneral,thestudysuggeststhatsomemechanistically related compounds can be grouped on the basis of their transcriptomesignature.

2. Materialsandmethods

2.1. Cellcultureandneuraldifferentiation

ThereporterhEScelllineH9 Dll1(GFPunderDll1promoter) was provided by Mark Tomishima from the Memorial Sloan Kettering CancerCenter (MSKCC,NY,USA).Importofthe cells and all experiments were carried out according to German legislation under license 1710 79 1 4 27 of the Robert Koch Institute (Berlin, Germany). The cell line was maintained on inactivated murinefibroblasts in DMEM/F12 medium supple mented with 20% serum replacement, HEPES (1M Gibco),

L gluthamine (Glutamax, Gibco), non essential amino acids (MEMNEAA,Gibco),

b

Mercaptoethanol(Gibco)andfibroblast growthfactor2(10ng/ml,Invitrogen).

H9hESCdifferentiationintoneuralcrestcellwasinitiatedon MitomycinCtreatedmurinebonemarrowderivedstromalMS5 cellsandcontinuedasdescribedpreviouslyinandFig.1ALeeetal.

(2010)andZimmeretal.(2012).Frozenstocksofthreeinvitro amplifiedNCCfromthreeindependentdifferentiationswereused foralltheexperiments.Eachstockwascarefullyqualitycontrolled basedonpreviousdefinedparameterssuchaspurity,differentia tionpotentialandmigrationcapacity.

2.2. Cellmigrationanalysis

Cellmigrationwasassessedwithminorchangesasdescribedin Zimmer et al. (2012, 2014). On day 2 NCC were thawed and 50,000cells/cm²wereseededin48wellplatesandweregrownto amonolayer.Aconfluentlayerofcellsistypicallyreached2days afterplating.Oncethecellshavereached100%confluenceacell free gap (scratch) wascreated using a 20

m

l pipette tip. After scratching,themediumwaschangedtofreshmediacontainingthe testcompoundsfor48h.Immediatelyafterscratchingthewidthof thecellfree areawasdeterminedin a controlplate.After 48h cytotoxicitywasanalyzedusingtheresazurinreductionassayas describedpreviously(Zimmeretal.,2012).Inordertodetermine migratedcells,thenucleiwerestainedwiththeDNAdyeH 33342 andunbiasedimages(3 4)alongthescratchweretakenusingan OlympusIX81at 4magnification. The numberof cells in the region ofinterest(ROI)wascounted automaticallybyKonstanz informationminer(KNIME)workflow(Bertholdet al.,2008).

2.3. Chemicalexposureduringmigration

hESCderivedneuralcrestcellswereexposedtochemicalsinN2 medium containing(Lee etal., 2010)EGF (20ng/ml) and FGF2 (20ng/ml)for48h.Foradetailedlistofthechemicalsandtheir concentrationusedinthisstudyseeFig.2.Theconcentrationof eachchemical waschosenafterassessingitsgeneral toxicityin NCCafter48hofexposure.Eachcompoundwastestedoverawide concentrationrange(severallogs)usingtheresazurinreduction assay. For each individual toxicant, the highest determined non toxicconcentrationwasthenusedintheMINCassay.

57

2.4. Konstanzinformationminer(KNIME)workflow

Theconsecutiveanalysis,i.e.thedetectionandpositioningof thescratchfollowedbythecountingofthecellsinsideascratch works as follows: in order to detect the scratch after 48h of incubation, threewellswerescratchedand images ofthethree wells were taken directly after scratching. Prior observations allowed ustoassume,thatthescratchintersectsahorizontally centeredline paralleltotheyaxis.Startingfromahorizontally centered1pixelwideboxthealgorithmiterativelyextendsthebox totheleftbyacertainwidthSineachiteration.Incasetheamount of thecells inside the newly added box of width S exceeds a threshold predefined T, the growing process stops and starts growing totheright.In thisway,foreach untreatedimagethe scratchisidentifiedandlocated.Theminimumwidthofabox,i.e.

the minimum width of the detected scratches, is used for the

succeedinganalysis,whereaboxofthiswidthisshiftedoverthe imagesfromthelefttotheright.Aftereachshiftthenumberof cellsintheboxisdetermined.Thepositionofthebox,wherethe minimalnumberofcellsisfoundisassumedtobethelocationof thescratch. Thisprocedure is necessary as the location of the scratchmayvaryinitshorizontalposition.Fig.1Cillustratesthis process(Bertholdetal.,2008).

2.5. Quantitativereal timePCR(qPCR)

qPCRsampleswereobtainedaftercellsweretreatedfor48h withtheindicatedtoxicantsattheindicatedconcentrations.Cells werelysedinTriFast^TM(Peqlab,Germany).TotalRNAwasisolated accordingtothemanufacturer’sinstruction,andtranscribedinto cDNA using the iScript Kit from BioRad (iScript^TM Reverse Transcription Supermix for RT qPCR, BioRad). Quantitative

Fig.1.‘‘Migrationinhibitionofneuralcrestcell’’(MINC)assaycombinedwithautomaticanalysis.(A)Schematicillustrationofthedifferentiationprotocolofhuman embryonicstemcells(hESC)toneuralcrestcell(NCC).(B)TreatmentprotocoloftheMINCassay.(C)ImageanalysisbytheKNIMEworkflow.Onday0thescratchwas detectedandtheaveragewidthofthescratchwasdeterminedfromthreetestwells.Aftera2dayincubationperiodunbiasedpicturesweretaken.Thesoftwareautomatically redetectedthescratchinthepicturesbyusingtheinformationfromday0anddeterminedthepositionofthescratchandtherebytheregionofinterest(ROI).Thecellsinthe ROIweredetectedandcountedautomaticallybythesoftware.Finally,thenumberofcellsmigratedintotheROIwasdefinedasthetotalnumberofmigratedcells.

(D)ValidationoftheKNIMEimageanalysisvs.manualcounting:theMINCassaywasperformedintheabsenceandpresenceofeightcompoundsatdifferentconcentrations.

Thenumberofcellsmigratingintothescratchwasanalyzedeithermanually(asdescribedbyZimmeretal.(2012))orbyusingtheKNIMEanalysistool,basedonthesame recordedimages(n=85).

real timePCR(qPCR)wasperformedonaBioRadCFX96Thermal Light Cycler (Biorad, Mu¨nchen, Germany). In order to keep technicalvariabilitylow,theRNAofallsamplesofeachreplicate wasprepared,andtranscribedinto cDNAin oneworking day.

Also,qPCRofsamplesofonereplicatewasperformedinparallelto avoidbiasintroducedbythesequenceofprocessingandanalysis.

Forquantification,qPCRthresholdcycleswerenormalizedina firststeptothehousekeepinggeneGAPDH.qPCRdataoftreated sampleswerenormalizedbysubtractionoftheuntreatedcontrol

anddisplayedasfoldchangerelativetocontrols(2^{( deltadeltaC(t)}

method)(LivakandSchmittgen,2001),asdescribedindetailearlier (Zimmeretal.,2012,2014).Thesequencesofspecificprimersare giveninFig.S1.

2.6. Biostatisticsanddatadisplayalgorithms

TheRsoftwareforstatisticalcomputingversion3.1.0,wasused asmaintoolforthestatisticalanalysis.

Fig.2.Tableofallcompoundsusedforthisstudy.Thetableindicatestheconcentrationusedformigrationanalysisaswellasforgeneexpressionanalysisforeachcompound usedinthisstudy.Wechosetoolcompoundsknownfrompreviousmigrationstudiesaswellascompoundswithsimilarstructureormodeofaction.^aNominalconcentration usedinthisstudy;^bAssociationoffigurereferencetopaperreferencelistnumber;^cinin-houseNCCmigrationassay(Zimmeretal.,2012);^dNCCgiverisetoribcage (Hendersonetal.,1999);(1)(Grandjeanetal.,1997);(2)(GrandjeanandLandrigan,2006);(3)(Opleretal.,2008);(4)(Haetal.,2009);(5)(Tianetal.,2009);(6)(Kippleret al.,2012);(7)(Hamadanietal.,2011);(8)(Sahaetal.,2012);(9)(Eubigetal.,2010);(10)(Patandinetal.,1999);(11)(Valvietal.,2012);(12)(Laietal.,2002);(13)(Chenet al.,1992);(14)(Zarnetal.,2004);(15)(DiRenzoetal.,2007);(16)(Machera,1995);(17)(GiaviniandMenegola,2010);(18)(Krieger,2004);(19)(Deanetal.,2002);(20) (Arndtet al.,2005);(21)(Meadoret al.,2006);(22)(Rasalamet al.,2005);(23)(Menegolaetal.,2005b);(24)(Verbois,2006);(25)(Valenzuela-Fernandezet al.,2008).

59

Expression profiles and individual compounds were hierar chically clustered using Euclidiandistance and average linkage method.Additionally,Kmeansclusteringwasperformedinorder toidentifygroupsofgeneswithsimilarexpressionpatterns.The optimalclusteringparameterswereselectedusingwithingroups sumofsquares andresultedin threeclusters. Matricesof gene expressionvalueswerevisualizedasheatmapsproducedusingR packagepheatmap(Kolde,2012).Empiricalpvaluesforcorrela tions of geneexpressionof compoundpairs werederivedfrom 10,000permutationsofthedatasetsandratioingofthoseruns withbettercorrelationsthanthePearson’srvaluesindicatedvs.

thetotalnumberofruns.

Principal component analysis (PCA) was performed on the original data set of 17 compounds. The first three principal componentscovered78%ofthevarianceinthedata.Inorderto visualize expressiondata inthree dimensions,representingthe first threeprincipal components(PC),the‘Excel3Dscatterplot’

macrowasused.

Forfurtheranalysis,10compoundswiththreereplicateswere selectedfromtheoriginaldataset.Inordertoinvestigatefurther the separation of the selected compounds, linear discriminant analysis(LDA)wasperformed.LDA(Cheoketal.,2003;Rencher, 2003)assumespredeterminedgroups(incontrasttoPCA)andthus belongstotheclassofsupervisedtechniquestestingthevalidityof theassumption(=grouping).However,LDAcannotbeappliedon

‘ill definedproblems’,i.e.whenthenumberoffeaturesislarger thanthenumberofobservations(Rencher,2003).Therefore,we firstappliedPCAontheselectedpartoftheoriginaldatasetto reducedimensionality,asproposedinBelhumeuretal.(1997).The first 13 principalcomponentsthat covered95% ofthe variance wereselectedforfurtheranalysis.

Theleave one outcross validationmethod(LOOCV)wasused toevaluate theLDA modelperformance withrespecttomodel accuracy(AccLOOCV).TheRpackagecaret(Kuhnetal.,2008)was usedformodelbuildingandperformanceevaluation.Inorderto assessthequalityofseparationachievedbydiscriminantfunctions inthegivendataset,apermutationtestwasused.Thelabelsofthe individualsampleswererandomlypermuted1012timesandused for model building. The model accuracy (AccRP) was then calculated on each permutation step and used to calculate an empirical pvalue underthe assumptionof thenull hypothesis H0: Acc_LOOCV=AccRP and of the alternative hypothesis H1:

AccLOOCV>AccRP.

Thecalculationanddisplayoftoxicitycurvesweredoneusing GraphPadPrism5.0(GraphpadSoftware,LaJolla,USA).

3. Results

3.1. Unbiaseddataacquisitionforneuralcrestmigrationinhibitionby automatedimagingofscratchassays

Migrationofneuralcrestcells(NCCs)isakeybiologicalprocess invertebratedevelopmentandapotentialtargetofdevelopmental toxicants.Inlinewiththis,ithasbeenshownthatthemigrationof neural crestcell assay (MINC) (Zimmer et al., 2012)is able to identify inhibitorsof NCC migration in a sensitive and specific manner. The NCCsused forourassay arederived fromhuman embryonic stem cells (hESC),and stored as frozen stocks until immediateuseintheMINC(Fig.1A).Severalacceptancecriteria suchaspurity>98%andpropermigrationbehaviorweredefined asqualitymeasuresforeachofthebatchesused.Forthepresent study,wefollowedtheestablishedprotocolandprobedmigration capacityofNCCsinthetimewindowof48 96hafterthawingand plating(Fig.1B).

The main endpoint of the assay requires manual, time consumingcountingofmigratedcells,and anestimationofthe

original scratch width(Zimmer et al., 2012). To make thetest system more operator independent, and to automate the data acquisitionandprocessingpipeline,anautomateddigitalimaging basedanalysisofthescratchwasdevelopedusingtheopensource KNIME(Konstanz information miner) software(Berthold et al., 2008). A protocol was developed, which measures the initial scratchwidth(day0),thenusesanalgorithmtoidentifyandre locatethescratchafter48h, andtoquantifythemigratedcells automatically(Fig.1C).

To compare the new approach (automated counting using KNIME)tomanualcounting,cellmigrationwasquantifiedin84 testconditions(fiveindependentexperiments,ninecompounds atvariousconcentrations)usingbothquantificationmethods.The results on migration inhibition did not differ significantly betweenbothmethods(Fig. 1D).OurdatasuggestthatKNIME analysismaybeslightlymore conservative(smaller inhibition values)than manualcounting, butdefinitely doesnot lead to oversensitivity of the assay. Since the new KNIME based quantificationmethodresultedinreliableandunbiasedresults, itwasusedtotestalargersetofpotentialtoxicantsintheMINC (Fig.2,Suppl.Fig.2).

3.2. Inhibitionofneuralcrestcellmigrationbydifferentgroupsof environmentaltoxicants

TotestwhetherthenewsetupoftheMINCassaywasableto confirmpositivesknownfrompreviousstudies(e.g.VPA,MeHg), andtoidentifynewhits wecompiled differentgroupsof well known environmental toxicants (metals, pesticides and poly chlorinatedbiphenyls(PCBs))tobetestedintheMINC.Initially,we determined the appropriate test concentration for each new compoundandpositivecontrols.Theaimofthispilottestingwas touseaconcentrationwithnosignificantcelldeathandatleast 25% inhibitionof migration. Therefore weperformed theMINC assay with increasing concentrations of the chemical and determinedcell viability by resazurinreduction and migration inhibitionafter48hexposurebyournewtestsetup,asdescribed above(Suppl.Fig.3A).

Compoundsnotinhibitingmigrationatanon cytotoxicconcen trationwereexcludedfromfurthertesting.Wefeltthatthispre selection of test concentrations and compounds was usefulfor comparing different endpoints of interest (mRNA changes vs.

migrationinhibition)withinthisinvitrocellsystem.

After the preliminary tests, compounds were tested in observer blindedfashionon4 5differentcelllotsinindependent experiments.Non cytotoxicitywasconfirmedfortheknownMINC positivecontrolsvalproicacid(1mM)andMeHg(50nM),aswell asforthenewpotentialtoxicantsAs2O3(1

m

M),CdCl2(100nM), cyproconazole (200

m

M), rotenone (10nM), PCB 170 (0.5

m

M), PCB153 (5

m

M) PCB 180 (5

m

M) (Suppl. Fig. 4A, Fig. 2). We compiledacomparisonwithinvivodosesorinvitroconcentra tions found/usedin otherstudies ofthe samecompounds.The concentrationsusedhere byuswerein concordancewithsuch literaturedata(Suppl.Fig.S2).

Allcompounds(atthechosenconcentration)blockedmigra tionsignificantly acrossall testrounds.Treatment withMeHg reduced migrationto 632.3%, andVPA reduced migration to 496% (meansSEM, as all following data in this paragraph).

Inhibition by As2O3 was to 585.4% of control, for CdCl2 to 654.7%, for cyproconazole to 695.4%, for rotenone to 707.5%, forPCB 170 to 7310.1%, for PCB 153 to 695.5%, forPCB180to44%12%(Fig.3A;Suppl.Fig.4B).Inacontrolstudy, cells were also treated with the negative control compounds acetylsalicylicacid(250

m

M)andacetaminophen(250

m

M)(Kader eit et al., 2012). Migration was not impaired by the negative controls(Fig.3B).

3.3. Upregulationofthreeexemplarymigrationrelatedgenes(CAV1, ITGA4,MYLK)byvalproicacid

Asafirstapproachtoexplorea potentialcorrelationbetween generegulationandalteredmigration,weselectedthreetranscripts for RT qPCR analysis. The choice of these markers for this preliminaryanalysis was drivenby aneasily available in house quantificationmethod.Wedecidedtousethesignalingcomponent CAV1(caveolin) (Lentini et al., 2008; Park and Han, 2009), the adhesioncomponentITGA4(integrin)(Deakinetal.,2009;Liuetal., 1999;Testazand Duband,2001)andtheRhoKinase (Groysman etal.,2008)effectorMYLK(myosinlightchainkinase)(Miaoetal., 2010),astheyhavebeenusedinearlierstudies,butfordifferentcell types,andtheycontributetodifferentmigrationrelatedprocesses.

Theirdifferentialgeneexpressionaftertoxicanttreatment(same concentrationasformigrationdata)wasthencomparedforselected

representativesofeachtoxicantgroup(As2O3,cyproconazole,PCB 180) and two positive controls (VPA, MeHg). Assay conditions correspondedexactlytotheMINCassayasdescribedabove(48h exposure,sameconcentrationandculturecondition).

Wefoundthattheanti epilepticcompoundVPAup regulatedall threetranscripts,comparedtonegativecontrolsoruntreatedcells.

MeHgdidnotclearlyalterexpressionofthesethreegenes;As2O3

andcyproconazole,atconcentrationsinhibitingmigrationapproxi matelytothesameextent,down regulatedthethreegenesandPCB 180 showed a mixed response pattern (CAV1 and ITGA4 were slightlydown regulatedandMYLKwasnotregulated)(Fig.4).

3.4. InhibitionofNCCmigrationbydiverseHDACinhibitors

VPA,aclinicallyuseddrug,hadoriginallyonlybeenusedasvery robustlyfunctioningpositivecontrolwithnoparticularrelationship Fig.3.Inhibitionofneuralcrestcellmigrationbydifferentcompounds.Neuralcrestcellsweretreatedfor48hwiththeindicatedcompounds;migrationintoregionof interest(ROI)wasassessedbytheMINCassayandcomparedtoanuntreatedcontrolasdescribedinthemethodsection.TheMINCassaywasperformedusingtheindicated compoundsinnon-cytotoxicconcentration,whichweredeterminedbyviabilitytestingusingresazurinreduction.Forallindicatedcompoundsviabilitywas100%atthe chosenconcentration(seeSuppl.Fig.S4).(A)MigrationwasinhibitedbythefollowingcompoundsVPA(1000mM),MeHg(0.05mM),As2O3(1mM),CdCl2(0.1mM),rotenone (0.001mM),cyproconazole(200mM),PCB153(5mM),PCB180(5mM),PCB170(0.5mM),PCB138(0.5mM).(B)TheMINCassaywasperformedusingnegative(ASS, acetaminophen)andpositive(MeHg)controlcompoundsusingthefollowingconcentrations:acetylsalicylicacid(250mM;ASS),acetaminophen(250mM)andMeHg (0.05mM).MeHgwastestedalonginthissetofexperimentsasacceptancecontrolforgeneralassayfunction.(C)InhibitionofmigrationaftertreatmentwiththeHDAC inhibitorstubacin(0.1mM),SAHA(0.5mM)andTSA(0.01mM).*p0.05,**0.01***0.001.Dataaremeansof4–5independentexperimentsSEMcomparedtountreated control(setto100%).Dottedredlinesindicatethe100%leveloftheuntreatedcontrolforbettervisualorientation.

61

toenvironmental toxicants.However,we were intriguedbythe significant,andunidirectionalup regulationofthethreemigration related genesMYLK, ITGA4and CAV1.One ofthe many known modesofactionofVPAisinhibitionofhistonedeacetylases(HDAC), andHDACinhibitionoftenleadstoalteredtranscriptionalactivity.

Thereforewewonderedwhetherinhibitionofneuralcrestmigration is a generaltoxicologicalfeature of HDACinhibitors(HDACi) or specificforVPA.Threeadditional,andmorespecific,HDACi,namely trichostatinA(TSA),suberoylanilidehydroxamicacid(SAHA)and tubacinwerethereforetestedintheMINCassay.VPA,TSAandSAHA inhibitallclassI+IIHDACs(DietzandCasaccia,2010),whiletubacin isspecific forHDAC 6(Dingetal.,2014).TheexposuretoSAHA (500nM), TSA (10nM) and tubacin (100nM) lead to a strong decreaseinNCCmigrationto669%,585%and726%ofcontrol (Fig.3C).Fromthesedata,weconcludedthatHDACidorepresentanew groupoftoxicantswithpotentialadverseeffectsonneuralcrestduring development indeeddue toHDACinhibition and notdue toother off targetsofVPA.Thereforewewereinterestedinseeingwhetherthey showedauniformgeneregulatoryresponse.

3.5. Confirmationofup regulationpatternbyHDACinhibitors

The preliminarydataon gene expressionchanges suggested specific patterns of transcript changes for different toxicant groups.Toobtainmoreinformationontranscriptchangeswithin onegroupoftoxicants,andtoexpandthenumberoftranscripts studied,weextendedouranalysisto35migration relatedgenes (Fig.5)and alargersetofcompounds(Fig.2). Thegeneswere selected based on migration related functions and changes observed during migration as well as altered expression in responsetoavarietyoftoxicantsthataltermigration.However itisnotclearwhetherchangesinthesegenesplayrolesincausing migration,respondtomigrationevents,orweresimplycorrelated withcellmigration.Thestudynowincludedthepilotcompounds VPA,MeHg,As2O3,cyproconazole,andPCB180.Italsoincluded theadditionalmetalsCdCl₂,additionalPCBsPCB170andPCB153 (PCBs),aswellasmorepositivecontrolsfrompreviousstudies,i.e.

triadimenol,triadimefon (triazoles),Pb acetateandthimerosal (metals)(Zimmeretal.,2012).NCCswereincubatedfor48hwith non cytotoxic, migration inhibitory concentrations of these compounds, before mRNA was isolated and analyzed. As a negativecontrol,mannitolwasincluded.Althoughtheobserved regulations did not reach the 5% significance level on the individual gene level (after false discovery rate correction),

clusteringandcolor codinginaheatmapsuggestedco regulation patterns.Moreover,hierarchicalclusteringclearlyindicatedtwo majorgroupsofcompoundsaccordingtotheirgeneregulation pattern(Fig.6,Suppl.Fig.5).Thefirstgroupincludedthegeneral HDACinhibitorsVPA,TSAandSAHA,anditwascharacterizedby upregulation of >50% of theanalyzed transcripts. The second cluster consisted of two subgroups of compounds. The first containedcyproconazole,thePCBs170and180,As2O3andCdCl2, andwas characterized byan overall down regulationof tran scripts.Thesecondgroupcontainedtheremainingcompounds and showed very little transcriptional change. This analysis confirmed the unique properties of HDACi that distinguished themfromothermigrationinhibitors.Italsoclearlyshowedthat thesameapicalendpoint(impairedmigration)canberelatedto largelydifferentchangesofintracellularmarkers.

Tocomparethephenotypicresponse(migrationdeficit)andthe corresponding gene expression changes we chose three com pounds(TSA,As2O3andPCB180)andanalyzedtheconcentration dependencyofgeneregulationforasetof5 7geneschosenfrom Fig.6.GeneexpressionanalysisrevealedthatthemRNAresponse curvesaremostly monotonic.Thissuggeststhattheexpression dataarenotrandomnumbers(althoughchangesaresmall),but biologicallyconsistent.Thetoxicantconcentrationsusedin this studywererightintherangewhentranscriptsstartedchanging andthechangeinRNAexpressionscaledwithtoxicantconcentra tion.Thissuggeststhattranscriptchangesandalteredmigration maybebiologicallyconnected,despitethesmalloverallchanges (Suppl.Fig.3B).

Togetinitialinformation onco regulation ofthemigration relatedgenes,weperformedkmeansclusteringofthe35markers acrossalltestcompounds.Thisresultedinthreedistinctclusters (Suppl.Fig.6A).Forinstance,theinitialmarkersCAV1andMYLK co segregated in cluster 1, and this cluster very strongly distinguished general HDAC inhibitors from other compounds.

Cluster2wasenrichedingenessuchascalpainorPTPN,whose cognateproteinshaveinhibitoryrolesinthemigrationnetwork (Suppl.Fig.6B).Cluster3containedtheinitialmarkerITGA4and manysignaltransductioncomponents.Clearpathwaysorbiologi cal principles did not emerge from this analysis across all compounds. Analternative approach, i.e. analysis of the genes co regulatedby onegivengroupofcompounds(e.g.HDACi)by search for overrepresented KEGG pathwaysand gene ontology termsyieldedtheobviousmigrationprocesses,butnoinformation beyondthis.

Fig.4.GeneexpressionchangesofthemigrationrelatedgenesCAV1,ITGA4,MYLK.Neuralcrestcellswereincubatedfor48hwiththeindicatedcompoundsexactlyasdone fortheMINCassayusingthesameconcentrationsshowntoreliablyinhibitNCcellmigration:(Mannitol(negativecontrol;250mM),VPA(1000mM),MeHg(0.05mM),As2O3

(1mM),cyproconazole(200mM)andPCB180(5mM));mRNAwasisolatedfromNCCafter48htreatmentwiththetoxicantsandanalyzedbyRT-qPCRforCAV1(black), ITGA4(gray)andMYLK(white).Geneexpressionlevelsarepresentedrelativetountreatedcontrols,andnormalizedtoGAPDH.Dataaremeansofthreeindependent experimentsSEM.

3.6. Correlationofgeneexpressionprofilesofstructurallyand mechanisticallyrelatedcompounds

Clustering gives equal weight to both low information components (i.e. minor regulations) and to biologically more

meaningfulinformation(regulatedgenes).Therefore,weapplied principalcomponentanalysis(PCA)tothetranscriptdatasetas alternativeapproach.Wewereinterestedwhethercompoundsof thesamemechanisticgroup(HDACi)wouldclustertogether,and separatefromothergroupsoftoxicants. Thefirsttwo principal Fig.5.Cellmigrationregulationpathways.(A)Schematicrepresentationofselectedgenesandtheirfunctionalrelationshipamongeachotherandtomigration.Genestested ongeneexpressionlevelinthisstudyaredepictedingreen.Additionalgenes,showninwhite,aredisplayedforpathwayorientation.(B)Abbreviationsandfullnamesofthe genesanalyzedinthisstudy(Forinterpretationofthereferencestocolorinthisfigurelegend,thereaderisreferredtothewebversionofthisarticle.).

63

components(PC1/PC2)covered 68%of thetotalvariance inthe data,and PC3additional10%.Thereforeanalysiswaslimitedto the first three PC dimensions (Fig. 7A). The class I/II HDAC inhibitorsVPA,TSAandSAHAwereclearlyseparatedfromallother compounds,evenwhenonlyPC1andPC2wereconsidered.Onthe projectiontothe2Dplane,theHDAC 6inhibitorstubacinlayright betweenthegroupofHDACiandthenegativecontrolmannitol.

Anothergroupthatseparatedwellonthefirsttwodimensionswas thetriazolescyproconazole,triadimefonandtriadimenol,aswell asthesmallgroupoflightmetals (As2O3andCdCl2).Theheavy metalcompoundsPb acetate,MeHgandthimerosaldidnotclearly separatefromthecontrol.ThePCBsdidnotclearlyseparatefrom othergroups,buttheyclusteredtogetherasonegroup.Thus,the PCA appeared to showa higher power of separation, than the standardclusteringalgorithms.Itwasinterestingthatsomeclear

grouping emerged already with the use of a relatively small numberofmarkergenes.

ThedisadvantageofPCAis thatthedescriptivedata(visual inspection)are hardtoquantify,i.e. theapparentclusteringof groupsisdifficulttotestforstatisticalsignificance.Therefore,we tested, whether other quantitativeapproaches would support these findings. First, we explored whether gene expression profiles of related compounds showed better correlation than unrelatedcompounds. Thescatterplotsofgeneexpressionfor VPAvs.TSA,PCB153vs.PCB180,andtriadimefonvs.triadimenol suggestedthatthere isin facta goodcorrelationof transcript changes, even though the regulations of individual genes as suchweremoderateonly,andnoneofthesinglechangeswere significant at the 5% level after performing t tests and FDR corrections(Fig.7B).

Fig.6.Differentialgeneexpressionof35migrationrelatedgenesoverindividualcompoundtreatments.Neuralcrestcellswereexposedtotheindicatedcompoundsfor48h (equalconcentrationasinthepreviousMINCassay;indicatedinFig. 2).RNAwasisolatedandgeneexpressionwasassessedbyRT-qPCR.Geneexpressionlevelsareindicated relativetountreatedcontrolandthereferencegeneGAPDH.Expressionlevelswerescaledandcolor-coded(z-score).Darkredandbluecolorsindicatehighandlowlevelsof expressionrespectively.ExpressionprofileswerehierarchicallyclusteredusingEuclidiandistanceandaveragelinkage.Dataaremeansofthreeindependentexperiments (Forinterpretationofthereferencestocolorinthisfigurelegend,thereaderisreferredtothewebversionofthisarticle.).

Fig.7.Correlationofcompoundtreatmentsontheirgeneexpressionprofile.Neuralcrestcellswereexposedtotheindicatedcompoundsfor48h(equalconcentrationasin thepreviousMINCassay;indicatedinFig.2).RNAwasisolatedandgeneexpressionwasassessedbyqPCR;differentialexpressionwascalculatedbysubtractionofuntreated control.(A)BasedonthegeneexpressiondataobtainedbyqPCRanalysisof35migrationrelatedgenesprincipalcomponentanalysis(PCA)wasperformed.Firstthree componentscovercorrespondinglyPC1:52%,PC2:16.3%,PC3:9.6%ofthevarianceinthedata.Projectionsoftheindividualexperimentstothetwo-dimensionalplanesreflect theseparationbythecorrespondingPCs.(B)Comparisonofgeneexpressionprofilesofstructuralorfunctionalsimilarcompounds(TSA/VPA,PCB180/153,triadimenol/

triadimefon)byscatterplotcorrelation.(C)Pearsoncorrelationofthegeneexpressionprofilesofthesamegroupofcompounds(HDACinhibitors,triazoles,metals).

Thep-valuesweredeterminedusingapermutationtest:thefold-changevalues(n=35)associatedwithoneofthecompoundswereshuffled10,000timesamongthegenes.

Theempiricalp-valueindicatedshowsthefractionofreshufflingsforwhichacorrelationcoefficientlargerthantheoriginalcoefficientwasfound.

65

In order toquantify this, correlationcoefficients (r)and the significanceofcorrelation(empiricalpvalue)werecalculatedfor relevant compound couples (Fig. 7C). Using TSAas a reference compound of HDACi, VPA and SAHA showed highcorrelation, while tubacin correlated somewhat less. Within the triazoles, triadimefoncorrelatedsignificantlywithitsmetabolitetriadime nolandwithcyproconazole.MeHgcorrelatedwiththeveryrelated organomercurial thimerosalbutnot withPb acetate, As2O3 and CdCl2.PCB180correlatedwithPCB153,andtoasmallerextent (p=0.02)withPCB170.

3.7. RobustnessofgeneexpressionchangesandofHDACgroup clustering

As second statistical quantification approach following the qualitativegroupingobservedinthePCA,wechoselineardiscrimi nant analysis(LDA). This methodassumes alreadygivengroups (incontrasttoPCA),butallowsthequantificationofsignificantgroup differences. LikePCA,itleadstodimensionreductionontolinear vectors(lineardiscriminants).Theseseparationscanthensubse quentlybeusedforstatisticaltesting.

Threereplicatesof10compounds(VPA,SAHA,TSA,triadime fon, tubacin, CdCl2, mannitol, cyproconazole, triadimenol and As2O3), i.e. 30 data samples (each based on 35 transcript measurements by PCR) were used for the analysis, and the replicatesofagivencompoundwereconsideredtoformagroup.

TheLDAmodelwasthenconstructedasdescribedinmethods.The model accuracy (AccLOOCV) determined by leave one out cross validation (LOOCV) was 63% with a 95% confidence interval rangingfrom44%to80%.Inasecondstep,apermutationanalysis (altogether1012permutations)wasusedtocalculateanempirical pvaluefortheclassificationaccuracy.Apvalueof210 ¹²was obtained, suggesting that the accuracy of the LDA model is significant.

Alltoxicantreplicatesclusteredverycloselytooneanother,and many compoundsseparated fromoneanother, whenthe first3 lineardiscriminants(LD)wereconsidered(Fig.8).Visualinspection showsthatthefirstLDseparatedfivecompounds(VPA,SAHA,TSA, triadimenol, and As₂O₃)clearly fromthe others.The second LD separatedtriadimefonandthethirdonewasnotsufficientalonefor separation of additional compounds. Performance of the LDA classifierwasexaminedbyitsabilitytocorrectlyclassifyindividual experimentsintothecompoundgroups.For7outof10compounds 2ofthethreereplicateswerecorrectlyassigned(Suppl.FigS7).

Thisfinal analysisconfirmed,that theintra compoundvariation (replicates) was significantly lower than the inter compound variation,atleastforthe HDACiande.g.triadimenol andAs2O3. The resulting findings are in line with our suggestion that the compoundspecifictranscriptsignatureswerestatisticallysignifi cant,eventhoughindividualgenevariationswerenot.

4. Discussion

TheMINCassayhasbeenestablishedinitiallytoprovideatest systemforneuralcrestfunctionbasedonhumancells(Zimmer etal.,2012).Thisassaywasthenusedinafeasibilitystudyforanin vitrotestbatteryintheareaofdevelopmentaltoxicology(Zimmer etal.,2014).Inthepresentstudy,thetestsystemwasdeveloped further for general applicability by exchanging the manual evaluationofmigratedcellsforanewunbiasedcomputer based analysismethod.Moreover,expressionlevelsofaboutthreedozen migration relatedgeneswerequantifiedasadditionalendpoint.

ThisdatasetwasgeneratedwithintheMINCtestsystemtoexplore whethertheadditionalbiologicalinformationlayerwouldallow grouping and sub classification of toxicants that inhibit NCC migration. We found that some groups of chemicals with

supposedlyrelatedmodesofactiondidclustertogeneexpression signatures.Forinstance,statisticalapproachesbasedontheoverall patternofalteredgeneexpressionshowedthatcompounds(VPA, TSAandSAHA)belongingtotheclassofHDACiclearlyclustered together.AlsoPCBsortriazolestendedtoformclusters,although notassharplydefinedasHDACi.Thus,incorporationofarelatively small number of mechanism based endpoints into the cell function basedMINC assayallowedfor groupingof compounds along their toxic mode of action. This finding suggests that incorporationofgeneexpressionpatternsasadditionalendpoints inphenotypic/functionalcellbasedassayscansupportbiological read across.Byusingsuchapproachesonthebackgroundofdata onmanywell characterizedtoxicantsinthesameassay,additional toxicologicalinformationcanbe derivedfor newtoxicants that wouldnotbeavailablefromtheprimaryendpointofthetests.

Cellmigrationisakeyprocessduringdevelopment,aswellasin variousdiseases(Franzetal.,2002).Migrationofneuralcrestcells fromtheneuraltubetotheirfinaldestinationisakeyeventduring thedevelopmentofmanytissues,includingtheperipheralnervous system, the heart and facial bones (Dupin and Sommer, 2012).

Disturbanceofsuchmigrationeventscanleadtoseveremalforma tionsordiseasesincludingHirschsprungdisease,TeratologyofFallot or DiGeorge syndrome in humans (Keyte and Hutson, 2012;

Menendez et al., 2013). Impairment of NC development and function after exposure to a variety of chemicals, including pesticides,planarPCBsoranticonvulsantdrugshasbeenobserved indifferentvertebrate testsystems(DiRenzoetal.,2007;Fuller etal.,2002;Grimesetal.,2008;Menegolaetal.,2005a;Papisetal., 2006). Suchanimal basedtestsystemsare verylimitedintheir throughput.Therefore,thedataprovidedoftencompareonlyfew compoundsanddoses.Testsystemsabletoexplorealargersetof compoundsincludinggroupsofstructurallyand/ormechanistically relatedcompoundsusefulforQSARorread acrossapproachesare rare.Anadditionallimitationformechanisticstudiescanbethata particularsystemevaluatesanintegratedendpointaffectedbycell Fig.8.Lineardiscriminantanalysis(LDA)tovisualizeseparationofgeneexpression profilesduetotoxicanttreatment.Thelineardiscriminantanalysis(LDA)describes theseparationofthegeneexpressionprofilesofnineselectedcompoundsplus mannitolasnegativecontrol(black).Thevaluesofthediscriminantfunctionsfor each individual experiment are shown onthe axes. Depictedare thesingle experiments(n=3)forthe10selectedtreatments.

migration,celldivisionandcelldifferentiationat thesametime (Moorsetal.,2007).Sincemechanisticconclusionsaredifficultor impossible,ifseveraldifferentbiologicalprocessescontributetothe overall readout, we have shown earlier that the MINC assay faithfullyassessescellmigrationindependentofdifferentiationand celldivision(Zimmeretal.,2012).Theincorporationoftheunbiased KNIMEpluginforquantificationofmigrationdataintheMINCnow furthersupportstheevaluationoftoxicanteffectsonmigrationonly.

TheidentificationofgenesthatwouldbespecificforNCcell migrationisnotaneasytask.Themechanismofcellmigrationare very well described in various immune cells, cancer cells (Lauffenburger and Horwitz, 1996), NCC, and central neural precursors (Gong, 2014; Hippenmeyer, 2014; Kuriyama and Mayor,2008;Kwanetal.,2012;Liu,2011;MayorandTheveneau, 2014).Comparisonofthedifferentcelltypesshowsthatthecore migrationmachineryisevolutionary conservedandsharedbyall cells,whiletheintegrationofsignals,thecontributionofvarious pathways,andthecombinationoffinalcellularprocessesmight differdramaticallyfromcell tocell,and evenfromstimulus to stimulus.ThereforethereactionofNCCtotoxicantsdiffersfrom thatofcancer cellsorevenfromthatofrelatedcentralnervous systemprecursors(Zimmeretal.,2012),eventhoughtheoverall geneticcontrolmachineryis mostlythesame.Many ‘migration genes’arealso involved in entirelydifferentcellularprocesses.

Thereforegeneontologygroups(GOs)formigrationcontainabout 1600genes.Manyofthemarehoweveronlyinvolvedinmigration inveryspecificsituations,e.g.afteractivationofacertainreceptor onaspecializedimmunecell.Forourstudy,weselectedabout70 ofthekeymachinerygenesbychoosingfewrepresentativesforthe majorprocessescontrollingcellmigration suchascytoskeleton rearrangements(RAC1,RHOA,ROCK),celladhesion(ITGA4,ITGB1, ITGB3),andtissueinvasion(MMP).Afterexaminingtheexpression levelsinNCC wecompileda finallistofrobustlyexpressed 35 genes,whichweusedfortheanalysis.

The usefulness of our approach wasmost evident fromthe confirmationofHDACiasneuralcresttoxicants.Thestartingpoint ofthestudywastheuseofvalproicacid(VPA),asconvenientstable and water soluble positive control, as we found previously (Zimmer et al., 2012) that it strongly inhibits NCC migration.

Thedrugisusedasantiepilepticandtodampenbipolardisease.

Theclinicalmodeofaction issupposed toinvolveinhibition of gamma aminobutyricaciddegradation,but it may alsoinvolve modulationofionchannels,wnt signalingandalteredphosphoi nositolturnover(Rosenberg,2007;TerbachandWilliams,2009;

Williamsetal.,2002).Moreover, VPAisa well knowndevelop mental neurotoxicant that induces severe malformations when takenduringpregnancy(Ornoy,2009).Whenourpilotexperiment indicatedgeneralgeneup regulationbyVPA,weconsideredthat histonedeacetylaseinhibitionmightplayaroleasrelevanttarget of VPA in the MINC. VPA is known to broadly inhibit the de acetylationofhistones,andtherefore,itsalternativeuseincancer therapyhasbeenconsidered.The increasedacetylationstateof histonesisknowntomakeDNAmoreaccessibleandthischromatin changepromotesrelativelyunspecificgeneactivation.Totestthe relevance of HDAC as target in NCC migration, additional compoundsweretestedthataremoreselectiveandpotentHDACi, andthatlackeffectsonotherVPAtargets.Theseallwerefoundto behits intheMINC,andtheytriggeredsimilargeneexpression signaturesasVPA.ThismakesitlikelythatVPAactedintheassay mainlyasHDACi.Interestingly,variousHDACihavebeenshownto inhibitinvasivemotilityofcancercells(Eyu¨pogluetal.,2005;Liu etal.,2003).Thecompounddoesnotseemtoaffectatargetinthe coremigrationmachinery,sinceitonlyaffectsmigrationinsome tumorcells,andifreducesmigrationofsomeimmunecells,butnot others(Brogdonetal.,2007).Ourdataobtainedwithtubacin,an HDACithatspecificallyinhibitstheisoformHDAC6(Haggartyetal.,

2003),suggestthatHDAC6inhibitionmayplayaroleintheMINC.

Thisisoformismainlylocatedinthecytosol,anditregulatesthe acetylation state of tubulin (Hubbert et al., 2002), which is important for cell motility (Tran et al., 2007). Our clustering analysissuggestedthatVPA,TSAandSAHAaremuchcloserrelated than this groupis related totubacin. Thisexample showsthat additional background information (e.g. knowledge on enzyme inhibition pattern)canbeusefultofurtherrefine thebiological read acrossofourgeneexpressionsignature.

Thefinding thatthe twolight metal salts,CdCl2 and As2O3, clearly separated from the heavy metal compounds MeHg, thimerosalandPb acetateintriguedus.Indeed,CdCl2andAs2O3

havebeenshownearliertoinfluencethegloballevelofDNAand histonemethylation,andtherebystronglyaffectinggeneexpres sion (Cui et al., 2006; Zhou et al., 2012). On the other hand, especially the mercurialcompounds have been shown tohave relativelylittleinfluenceongeneexpressionalsoinotherhuman stemcell basedtestsystems(Krugetal.,2013).Theheavymetals mayexerttheirtoxiceffectdirectlyontheproteinlevelvia,e.g.

directenzymeinhibition.Forinstance,mercurydirectlybindsto freeproteinthiolgroups(Kandaetal.,2012;Rochaetal.,2012).For example, mercury can directly inhibit

g

secretase and thereby disrupt the notchpathway during neural development (Alattia etal.,2011).Directbiochemicaleffectsaspossiblemodeofaction for the heavy metals is further supported by the fact that the pesticiderotenoneclusteredinthesamegroup.Itiswellknown that rotenoneinhibits themitochondrial respiratory chain,and therebydepletesintracellularATP(Lietal.,2003)independentof geneexpressionchanges.

For all the other chemicals used, grouping based on gene expression wasless clearthan for HDACi.This was somewhat astonishing as, e.g. the groups of triazoles or PCBs contained chemicalsthatarestructurallycloselyrelatedandareexpectedto haveatleastrelatedmodesofaction.However,differentPCBsmay havedifferentaffinitiesforthesamereceptororhaveadditional targets,whichcouldexplainthedifferentgrouping.Moreover,the chosengenesetwasverylimitedandalargersetmayhaveshown similarities and differences more clearly. The grouping of chemicalsbasedongeneexpressionchanges isanewapproach andfewpublicationsareyetavailableonthistopic.Amongstthem, twostudiesexpandedtheembryonicstemcelltest (EST)byan additionaltranscriptomicsendpoint.Inthefirst,sixphtalateswere comparedtosixtriazoles,withthegroupsalreadyknownatthe onset of the study, and with whole transcriptome information provided(vanDarteletal.,2011).In thesecondstudyfromthe samelab(vanDarteletal.,2010),twoindividualcompoundswere comparedtoseewhether transcriptomedatawoulddistinguish theminanembryonicstemcellassay.Inbothstudies,thegroups tobedistinguishedweregivenbeforestatisticalanalysis,whileour study performed initial grouping without pre defining the differentcompoundclasses.

Theinclusionofahighernumberofgenesinourstrategymay eventuallyleadtotheidentificationofbiomarkersoftoxicity(BoT).

A BoT is defined as ‘‘a parameter that provides quantitative informationthatismechanisticallyrelevanttoandpredictiveofan adverse effect’’ (Blaauboer et al., 2012). Biomarkers play an important role especially for in vivo studies to get early and sensitive indications of liver or kidney injury (Dedeoglu et al., 2013;Schomakeretal.,2013).Moreover,theyareimportantforin vitrostudieswithcelldeathasprimaryendpoint.Insuchassays mechanisticbiomarkersallowtoobtainmoreinformationonthe modeofactionofacompoundthatcanbetranslatedtothehuman situation.ThisgoalislessimportantfortheMINC,asitalreadyuses a functional endpointrelevant to human toxicology.In sucha system the use of intracellular markers has slightly different priorities.Thetwoapplicationsare(i)mechanisticmarkersthat 67