A pan-European Infrastructure for Quality in Nanomaterials Safety Testing

NUID UCD | NHM | IOM | JRC | BFR | KIT | FUNDP | IST | UNIVLEEDS | NILU | HMGU | LMU | CIC | UU | ICN DLO/RIKILT | WU | DGUV | TAU | VITO | SMU | TCD | FIOH | UOE | HW | CNRS | INERIS | UBRM | RIVM

QuantitativeStructureActivityRelationship(QSAR)models

27.03.2013

Xue Z. WangCeydaOkselUniversity ofLeeds

Theory ofQSAR NewDataMining Methods NanomaterialCharacterization Nanotoxicity Application ofQSARmodelinnanotoxicitymodeling

2QualityNanoTrainingSchool

Purpose ofToday

Part ITheory ofQSAR

3QualityNanoTrainingSchool

QSARHistory

CorwinHermanHansch(1918 2011)

Thefatheroftheconceptofquantitativestructureactivity relationship (QSAR),thequantitativecorrelationofthephysicochemical

propertiesofmoleculeswiththeirbiologicalactivities[5]

4

1865

Crum BrownandFraser[1]

expressed theideathat there

was a relationshipbetweenactivityandchemicalstructure.

1900

Meyer[3] andOverton[4]found

linearrelationshipsbetweenthe

toxicityoforganiccompoundsandtheir lipophilicity.

1969

Hansch[5] publishedafreeenergyrelatedmodelto correlatebiologicalactivities

withphysicochemical

Properties.

1893

Richet[2]correlatedtoxicities of

simple organicmoleculeswiththeirsolubilityin

water.

[1]A.CrumBrown,T.R.Fraser,Trans.R.Soc.Edinb.25(19681969) 257. [2] M.C.Richet,Compt.Rend.Soc.Biol.45(1893)775.[3]H.Meyer,Arch.Exp.Pathol.Pharmakol.42(1899)109. [4] E.Overton,Studien uber dirNarkose,GustavFischer,Jena,1901[5] Hansch,C.(1969)AquantitativeapproachtobiochemicalstructureactivityrelationshipsAccountsofChemicalResearch, 2, 232239.

QualityNanoTrainingSchool

MolecularStructure

Physicochemicalor Biological Property

TheVarianceIn

MathematicalDependencies

Toxicity/Ecotoxicity Prediction

BulkFormNanoparticles(widely used)(notadopted)

A QSAR is a statistical modelthat relates a set of structuraldescriptors of a chemicalcompound to its biologicalactivity.

5

QSAR(QuantitativeStructureActivityRelationship)

What isQSAR?

[1]T.Puzyn etal.(eds.),RecentAdvancesinQSARStudies,103125. DOI10.1007/9781402097836_4,CSpringerScience+Business MediaB.V.2010QualityNanoTrainingSchool

The presenceofparticular characteristics increasethe propability that the compund istoxic

QSAR(QuantitativeStructureActivityRelationship)nanoQSAR

QSARmodels are very useful incase ofthe classic chemicals buttheconcept ofnanoQSARisstill under development.

QSAR has been widely used to predict the toxicity of substances in bulkform (especially druglike compunds) but, up to date, QSAR studies for theprediction of nanoparticle toxicity have been rarely reported.

Successful QSAR studies What about Nanosize particles?

6QualityNanoTrainingSchool

The European system for new chemical management (REACH) promotesQSAR methods as an alternative way of toxicity testing.

INSILICO IN VITROINVIVO(computational) (incontainer) (onlivingorganism)

predict predict

All chemical substances need to betested interms oftheir toxicologicaland environmental properties before their useThere are several reasons to use QSARModels :very fast,often free,reducethenumberofanimalsusedinexperiments

validate

Themostimportantsourcesofinformation

Alternative,fastdevelopingapplications

Reduction inthe time,cost and animal testing

themostpromising

QSARquantitativestructureactivity

relationship[1]

7[1]A..Gajewicz,etal.,Advancingriskassessmentofengineerednanomaterials:Applicationofcomputationalapproaches,Adv.DrugDeliv.Rev.(2012),doi:10.1016/j.addr.2012.05.014[1]

Sourcesofinformation[1]Why dowe need QSARmodels?

QualityNanoTrainingSchool

8

Purpose ofQSARTo predict biologicalactivityand physicochemicalpropertiesbyrationalmeans

To comprehendandrationalizethemechanismsofactionwithinaseriesof chemicals

Savingsinthecostofproductdevelopment

Predictionscouldreducetherequirementforlengthyandexpensiveanimaltests.

Reductionofanimaltests,thusreducing animaluseandobviouslypainanddiscomforttoanimals

Otherareasofpromotinggreenandgreenerchemistrytoincreaseefficiencyand eliminatewaste

QualityNanoTrainingSchool

9

Requirements for QSAR

How manycompounds arerequiredtodevelopaQSAR?

There isnodirectandsimple answer

asmany aspossible

Toprovidesomeguide,itis widelyacceptedthatbetweenfiveandtencompoundsarerequiredforeverydescriptor inaQSAR[1]

[1]Aptula AO,RobertsDW(2006)Mechanisticapplicabilitydomainsfornonanimalbasedprediction oftoxicologicalendpoints:Generalprinciplesandapplicationtoreactivetoxicity.Chem ResTox.19:10971105

1.Setof 2. Setofmolecular 3.Measured biological 4.Statisticalmolecules descriptors activity or property methods

Toxicity EndpointsMolecular Modelling

e.g. Neural NetworksPLS, Expert Systems

DESCRIPTORSPhyscochemical, biological, structural

QSARs

Daphnia magna EC50

CancinogenicityMutagenicity

Rat oral LD50Mouse inhalation

LC50Skin sensitisation

Eye irritancy

QualityNanoTrainingSchool

Applications ofQSAR

10

There are a large number of applications of these models withinindustry, academia and governmental (regulatory) agencies.

The estimation ofphysicochemical properties,biologicalactivities and understanding the physicochemical featuresbehind abiological response indrug desing.

Therationaldesignofnumerousotherproductssuchassurfaceactiveagents, perfumes,dyes,andfinechemicals.

Theidentificationofhazardouscompoundsatearlystagesofproductdevelopment,thepredictionoftoxicitytohumans and environment.

The prediction offate ofmolecules which are releasedinto the environment.

Thepredictionofavarietyofphysicochemicalpropertiesofmolecules.

QualityNanoTrainingSchool

11

What isrequired for agood QSARModel?

The QSARmodelshould meet therequirements ofthe OECDprinciples [1]:

[1]OECD,GuidanceDocumentonthe Validationof(Quantitative)Structure ActivityRelationshipsModels, Organisation forEconomicCooperation andDevelopment(2007).

1 adefinedendpoint;

2 anunambiguousalgorithm;

3 adefineddomainofapplicability;

4 appropriatemeasuresofgoodnessoffit,robustnessandpredictivity;

5 amechanisticinterpretationifpossible.

UninformativedescriptorsPoordescriptorselectionandchancecorrelationsModelling complex,nonlinearstructurepropertyrelationshipswithlinearmodelsIncorrectlyvalidatingQSPRmodelsNotunderstandingthedomainofapplicabilityofmodels

Common QSARModellingErrors

QualityNanoTrainingSchool

12

Methods

Molecular Structure

Representation

Molecular Descriptors

Predictive Model

ModelValidation

Applicability DomainofQSAR

Types ofMolecular Descriptors(topological,geometric,electronic and hybrid)

1DDescriptors 2DDescriptors 3DDescriptors

Main Steps

QSARModelling process consists of5main steps.

QualityNanoTrainingSchool

Begins with the selection ofmolecules to beusedSelection ofdescriptors;numericalrepresenter ofmolecular features (e.g.numberofcarbon)Original descriptor pool must bereduced insizeModelbuildingThe reliability ofthe modelshould betested

13

Molecular Descriptors[1]are numerical values that characterize properties ofmolecules.

cacalculated by

Mathematicalformulas

Computationalalgortihms

TheoreticalIndexes

Experimentalmeasurements

MolecularDescriptors

MolecularDescriptors

Different Theories(e.g.Quantumchemistry,

information theory,organic chemistry)

derivedfrom Scientific Fields (e.g.QSAR,

medicinalchemistry,drugdesign, toxicology,analytical

chemistry,Environmetrics, virtual

screening)

appliedin

statisticschemometrics

chemoinformatics

processedby

[1]T.Puzyn etal.(eds.),RecentAdvancesinQSARStudies,103125. DOI10.1007/9781402097836_4,CSpringerScience+Business MediaB.V.2010[2] Todeschini R,Consonni V(2000)Handbookofmoleculardescriptors.WileyVCH,Weinheim

More than 5000[2]

QualityNanoTrainingSchool

RegressionProblems Multiplelinearregression Partialleastsquares Feedforwardbackpropagationneuralnetwork GeneralregressionneuralnetworkClassificationProblems LinearDiscriminant analysis Logisticregression Decisiontree Knearestneighbor ProbabilisticNeuralNetwork SupportvectormachineRecentlydevelopedMethods Kernelpartialleastsquare Robustcontinuumregression Locallazyregression Fuzzyintervalnumberknearestneighbor Fastprojectionplaneclassifier

14

DataAnalysis

ModelValidation

DataCollection

MolecularDescriptor

[1] Ekins, S. (2007). Computational Toxicology: Risk Assessment for Pharmaceutical and Environmental Chemicals. New Jersey: Wiley-Interscience.

DataAnalysis Methods[1]

QualityNanoTrainingSchool

Different Statistical Methods have been used inQSARfor the extraction ofuseful informationfrom the data.

DataMiningMethodsarelativelynewfieldwhichsharesometechniqueswithtraditionaldataanalysisbutalsobringsalotofnewvisionsandtechniques.

Data:recordsofnumericaldata,symbols,images,documents

Knowledge:Rules:IF..THEN..CauseeffectrelationshipsDecisiontreesPatterns:abnormal,normaloperationPredictiveequations

Data Data Data Data

Information

Knowledge

Decision

Volume

Value

DM is the discovery of useful Informationand Knowledge from a large collection ofData, in order to help the Decisionmaking

15QualityNanoTrainingSchool

16

Clustering

Classification

Conceptual Clustering

Inductive learning

Data mining shares many techniques and tools with traditional dataanalysis (such as the ones listed below). But it emphasises the analysisof very large databases, and can deal with complicated cases whichcannot be handled by traditional methods.

Data mining research has also generated some new techniques thatwere not seen in traditional data analysis, such as Link Analysis

QualityNanoTrainingSchool

Summarisation

Regression

Case-based Learning

DataMiningMethods

17

x1 x2 x31 0 01 1 10 0 11 1 10 0 00 1 11 1 10 0 01 1 10 0 0

x1 x2 x3

x1x2

x3

Example :Tendatacases,threevariables,x1,x2,andx3 How canweautomaticallyuncoverthedependencyrelationshipbetweenthem?

Thisrelationshipmeans,x2 isdependent onx1,andx3isdependent onx2,butx1isindependentofx2andx3

Thisrelationshipmeans,x2isdependent onx1,x3isalso dependent onx1,butx2andx3areindependentofeachother

DATASET WHICHRELATIONSHIP?

QualityNanoTrainingSchool

NewDMTechniqueExample1:LinkAnalysis

18

x1 x2 x3 x4 y x1 x2 x3 x4 y55.33 1.72 54 1.66219 92.19 71.31 3.44 55 1.60325 90.5159.13 1.2 53 1.58399 92.74 72.3 4.02 55 1.66783 90.2457.39 1.42 55 1.61731 91.88 68.81 6.88 55 1.69836 91.0156.43 1.78 55 1.66228 92.8 66.61 2.31 52 1.77967 91.955.98 1.58 54 1.63195 92.56 63.66 2.99 52 1.81271 91.9256.16 2.12 56 1.68034 92.61 63.85 0.24 50 1.81485 92.1654.85 1.17 54 1.58206 92.33 67.25 0 53 1.72526 91.3652.83 1.5 58 1.54998 92.22 67.19 0 52 1.86782 92.16

x1,x2,x3aredifferentmeasuresoffeedcompositionx4isthelogofacombinationofoperatingconditionsy productqualitymeasurey=good if>92,=low ifbetween9192,=verylow if

19

Ifx1=[44.6,55.4],thenconditionalprobabilityofy=3is0.9

Ifx4=[1.8,2.3],thentheconditionalprobabilityofy =3is1.0

x1

y>92 x4

x1 Y>92

91

20

(1)Generationofapopulationofsolutions

(2) Repeatsteps(i) and(ii) untilthestopcriteriaaresatisfied:(i)calculatethefitnessfunctionvaluesforeachsolutioncandidate

(ii)performcrossoverandmutationtogeneratethenextgeneration

(3) thebestsolutioninallgenerations:thefinalsolution

s

.. .... ..

s

.. .... ..

s

.... ..

s

.. .... ..

s

.. .... ..

s

.. .... ....

s

.... ..

s

.. .... ..

Automaticgenerationof

decisiontreefromdata

This isanewapprooach

based onageneticprogramming

method

QualityNanoTrainingSchool

21

Data Set 1:Concentration lethal to 50% of the population, LC50, 1/Log(LC50), of vibrio fischeri, a biolumininescent bactorium

75 compounds 1069 molecular descriptors calculated by molecular modelling software DRAGON

Data Set 2:Concentration effecting 50% of the population, EC50 of algae chlorella vulgaris, by causing fluoresceindiacetate to disappear

80 compounds 1150 descriptors

Two Data Sets

QualityNanoTrainingSchool

New DM Technique Example

22

SampletreefromYAdapt

Class1(21/21)

Class1(8/8)

No

NoYes

Yes

Gearyautocorrelationlag4weightedbyatomicmass

0.007

Polarsurfacearea40.368

Class2(9/9)

NoYes

Hautocorrelationlag0weightedbySandersonelectronegativities2.111

Gearyautocorrelationlag4weightedbypolarizability

1.181

Class3(9/10)

Yes No

BrotoMoreauautocorrelationlag2weightedbypolarizability

4.085

Class2(5/6)

Yes No

Class3(6/6)

Atreeforthebacteriadatausingthree

classes.

Forbacteriadata:of 1069 descriptors,only5descriptorsareimportant! Usingthe5descriptors,thetree

modelcanpredict withthe accuracy>96.7%

QualityNanoTrainingSchool

YAdaptforQSARToxicityPrediction:DataSet1Amethod to classify toxicity endpoints into classes

23

A decision tree from generation 46 for the bacteria data, finding fourclasses from the continuous endpoint itself during induction.

Training accuracy: 93.3%Test accuracy: 80.0%

NoYes

Yes

NoYes

Yes No

No NoYes

Class1(13/13)

3DMoRSEsignal15weighted18weightedbySandersonelectronegativites0.644

Gearyautocorrelationlag4weightedbyatomicmass0.007

Polarsurfacearea40.368

Hautocorrelationlag0weightedbySandersonelectronegativities2.111

Class2(5/7)

Class3(7/9)

Class4(6/6)

Class1(16/16)

Class2(9/9)

Hautocorrelationlag0weightedbySandersonelectronegativities2.394

Class1: Log(1/EC50)4.0829casesClass2: 4.08

24

NoYes

Class2(14/15)

Class1(16/16)

No Yes

Class3(6/8)

NoYes

Yes No

Selfreturningwalkcountorder83.798

Hautocorrelationoflag2weightedbySandersonelectronegativities

0.401

Molecularmultiplepathcountorder392.813

Solvationconnectivityindex 2.949

Class4(6/7)

NoYes

2ndcomponentsymmetrydirectionalWHIMindex

weightedbyvanderWaalsvolume0.367

Class3(9/10)

Class4(8/8)

2nd dataset algaedata

Ofthe1150descriptors,only5descriptorsareimportant!

Training:92.2%Test:81.3%

QualityNanoTrainingSchool

YAdaptforQSARToxicityPrediction :DataSet2

25

ResultsPresentationGraphsTablesASCIIfiles

DiscoveryValidationStatisticalsignificanceResultsfortrainingandtestsets

DataMiningToolboxRegression PCAandICAART2networksKohonennetworksKnearestneighbourFuzzycmeansDecisiontreesandrulesFeedforwardneuralnetworks(FFNN)SummarystatisticsVisualisationMixtureQSARmodels

DataimportExcelASCIIFilesDatabaseXML

DataPreprocessingScalingMissingvaluesOutlieridentificationFeatureextraction

Descriptorcalculation

QualityNanoTrainingSchool

Different dataminingtechniques canbe

combined intoasinglesystem

IntegratedProcessDataMiningEnvironment

The statistical fitofmodelR2

StandartError ofPredicton (S)

Crossvalidation (Q2) Kfold Cros Validation Leaveoneout

Cross Validation

26

ModelValidation

QSARModelValidation

ModelApplicability Modelvalidationisrequiredtoensure

modelreliability (quality ofthe model)Ensures thatthemodelisnotdueto

chancefactors!Model

Interpretation

Cross validation:dividing adatasample into subsets,performing the analysis onasingle subset,using other subsets to confirm and validate the initial analysis

QualityNanoTrainingSchool

There isnoagreed method for the validation ofQSARModels

The statistical fitofmodelisusually performed

27

QSARWorkflow[1]

[1] Tropsha,A.,Gramatica,P.,Gombar,V. (2003) Theimportanceofbeingearnest:ValidationistheAbsoluteEssentialforSuccessfulApplicationandInterpretationofQSPRModels.Quant.Struct.Act.Relat.Comb.Sci.22,6977. QualityNanoTrainingSchool

28

QSARModelApplicability Domain

... .

.........

Descriptor 1

...

..

.. ..

..TRAININGSETDescriptor 2

Inside ApplicabilityDomain

will bepredicted

Outside ApplicabilityDomain

will notbepredicted

Notevenasignificant andvalidatedQSARmodelcanbeused for the entireuniverse of

chemicals[1]!

[1] Gramatica P. (2007)PrinciplesofQSARmodelsvalidation:internalandexternal.QSARandCombinatorialScience; 26:694701.

Modelapplicabilityallowsustodecidewhetherthemodelwill beusefulfornewdata (correct use of

the model).

Finding the boundaries ofthemodelto see how well itwill work

for other compounds.

QualityNanoTrainingSchool

29

QSARTools for Toxicity Prediction1. TOPKAT(byKomputer AssistedTechnology):

ItusesQSARstogivenumericpredictions,basedonseveral descriptors

2. HazardExpert (by CompuDrug ):HazardExpert,producedbyCompuDrug,isarulebasedsystem.Therulesusetoxicophores,orstructuralalertsderivedfromQSARs

3. CASE: It canderive QSARmodelfrom usersdata.

4. ECOSAR:It usesQSARstopredicttheaquatictoxicityofchemicalsforavarietyoforganisms

5. OASIS: ItuseshierarchicalrulesbasedonQSARs,whichcanbeuserderived.

6.DEREK:DeductiveEstimationofRiskfromExistingKnowledge(DEREK)ismarketedbyLHASALtd.It isanexpertknowledgebasesystemthatpredictswhetherachemicalistoxic

QualityNanoTrainingSchool

Each ofthese QSARtools have their own advantages and weaknesses for toxicitypredictions

There are also several available softwarepackages for descriptor calculations (e.g.DRAGON)

30

Challenges for QSAR[1]

[1]Puzyn,T.,Leszczynski,J.(2012) TowardsEfficientDesigningofSafeNanomaterials:InnovativeMergeofComputational ApproachesandExperimentalTechniques,TheRoyalSocietyofChemistry.

1

If there isameasurement error inthe experimental data,itisvery likelythat false correlations may arise

2

Ifthetrainingdatasetisnotlargeenough,thedatacollectedmaynotreflectthecompletepropertyspace.Consequently,manyQSARresultscannotbeusedtoconfidentlypredictthemostlikelycompoundsofthebestactivity.

3

Therearemanysuccessfulapplicationsofthismethodbut onecannotexpecttheQSARapproachtoworkwellallthetime. QSAR Modelsareeasytobuildbutalsoveryeasytogetwrong.

QualityNanoTrainingSchool

Part IINanotoxicity and

NanomaterialCharacterization31QualityNanoTrainingSchool

Nostandardizedorvalidatedmethodsfornanotoxicity

testing

Available toxicitydatafor NPs areinconsistent and

confusing

Blocks thedevelopmentof

ENPsriskreductionstrategies[2]

NewTechnology

NewConcerns[1]

NewBenefits

Safetytohumanhealthand environment Suitability ofriskmanagementstrategies

Macroconcernswithnanoworld !

32

[1] Chaudhry, Q. et al. (2010). Knowns, Unknowns, and Unknown Unknowns . RSC Nanoscience & Nanotechnology (p. 201-218).Royal Society of Chemistry.[2]Tran, L., & Chaudhry, Q. (2010). Engineered Nanoparticles and Food: An Assessment of Exposure and Hazard. RSC Nanoscience & Nanotechnology (p 120-134). Royal Society of Chemistry.

Macroconcernswithnanoworld !

QualityNanoTrainingSchool

Uncertainities atthe early stages ofanew technology bring new concerns!

More than 800nanoproducts

33

Riskassesment for NMS[1]

1.HazardIdentification

2.Doseresponse

relationship

3.Exposureassessment

4.Riskassessment

Whatdangerstohuman healthmaybasicallyarisefromtheharmful agent ?

Whatisthenatureandmagnitude ofthe

riskandhowaccuratelycanitbe

estimated?

Towhatextentistherelevant

populationgroupexposed totheharmful agent?

Whatquantitativeconnectionsexistbetweenthedose

andtheextentoftheexpected effect?

OECD (2003): Environment Directorate. Description of Selected Key Generic Terms Used in Chemical / Hazard Assessment. OECD Series on Testing and Assessment Number 44. ENV/JM/MONO(2003)15. Paris, 30-Oct-2003.

Riskassesment for NMS[1]

Potential safety issues ofNMs should beaddressed atthe same timeasthetech.isdeveloping

QualityNanoTrainingSchool

ENMs are already used inanumber ofcommercial applications.

34

Hazard Identification

Doseresponse evaluationEvaluation oftoxicity

involves determiningtheconditionsorlevelsatwhichthepresenceofacontaminant maytriggerabiologicalresponsefromthebody

Different amount ofCNTdust

At level of0.1mg/m3

Athigher level of0.5mg/m3

Noeffects wereobserved

Sideeffects weredetected in the

lungNaidu, B. David (2009). Biotechnology and Nanotechnology - Regulation under Environmental, Health, and Safety Laws.. Oxford University Press.

Doseresponse relationship

QualityNanoTrainingSchool

Theevaluationoftoxicityincludes twosteps:

35

Theonlyreasonable waytoobtain toxicity information for thelargenumberofnanoparticleswithout testing everysingleone istorelate

thephysicochemicalcharacteristicsofNPs withtheirtoxicityinaSAR

(StructureActivityRelationship)orQSARmodel.

Nanomaterialcharacterization isone ofthe first steps intoxicity test.However,itisavery challenging issue!!!

QualityNanoTrainingSchool

Particle Characterization

NanoparticleStructureProperties

ParticleSizeSurfaceAreaMorphology

StateofDispersion,

Toxicity

Sizeandshape Sizedistribution Agglomeration state Porosity Structuredependentelect. Electronicproperties Characterisation:

Inwhatmedium?

POSSIBLE factors[1]inducingtoxicityof

chemicals

Surfacearea Surfacechemistry Surfacecharge Crystalstructure Composition Configuration Coating Aggregationstate Metalcontent

When amaterial isinananosized form,its properties become different from thesame material inabulk form

Fullerenes Carbonbased Different and CarbonNTs Nanomaterials Exceptional Graphene Properties

36[1] Puzyn, T., Leszczynska, D. & Leszczynski, J. Toward the development of Nano-QSARs: advances and challenges. Small 5, 24942509 (2009).

Possible toxicityrelated Physicochemical Properties

QualityNanoTrainingSchool

Nocertain information!!!Recent studies showed that toxicity ofNPs canberelated to:

moreresearchisrequiredtoreachaclearunderstanding

more research is required to reach a clear understanding of the

Shapeandaggregationstate

Composition,purityandsurfacechemistry

Surfacecharge

Crystalstructure

Particlenumberandsizedistribution

SurfaceArea

HighResolutionMicroscopy

SpectroscopyandChromatographyMethods

ZetaPotentialAnalysis

XRayDiffraction

SeveralTechniques(e.g.SEM,TEM,DLS)

BrunauerEmmettTellerAdsorption

Several different methods and instruments are used to determine the properties onNMs

37

NanomaterialCharacterization

QualityNanoTrainingSchoolThere isnosingle instrument that isright tool for every test.There are infact more than 400

different techniques for particle counting and characterization

38

SomeNPs

Have atendency toformclusters (e.g.aluminium oxide)

inanagglomeration state,someNPsbehavelikelargerparticlesinthemedia

Toxicitymayalsodependsonthesizeofagglomerate

NOTonoriginalparticlesizeitself!!!

Hussain, S. e. (2008). Can silver nanoparticles be useful as potential biological labels? Nanotechnology 19 .

If the NPs aggregateitisimportant toidentifyhowthesechangesaffecttoxicity

Howtocharacterizeaggregation?Textureanalysis?

Whathaven'tweconsideredyet ?

39

ChallengesforQSAR

NPtype

Size(nm)

Doses(mg/ml) Cell Method Result

Ag 720 0.850Humanfibrosorcoma andskincarcinoma

XTTassay Nontoxic Araro etal.(2008)[1]

Ag

40

Challenges for nanoQSAR[1]

[1]T.Puzyn etal.(eds.),RecentAdvancesinQSARStudies,103125. DOI10.1007/9781402097836_4,CSpringerScience+Business MediaB.V.2010

1 Lackofhighqualityexperimentaldata

2 Lackofconceptualframeworksforgrouping NPs accordingtomodeofphysicochemicalpropertiesandtoxicaction

3 Lack ofsufficient moleculardescriptors abletoexpressspecificityofnanostructure

4 Lackofrationalmodelingprocedurestoscreenlargenumbersofstructurallydiversifiednanoparticles.

5 Limited knowledgeontheinteractionsbetweenNPs andbiologicalsystems

QualityNanoTrainingSchool

NMs are problematic for QSARmodellers because ofthe several reasons:

Part IIICase Study

despite all limitations41QualityNanoTrainingSchool

Nanotxicology 2012,47September,Beijing

The aim of this study was to test the hypothesis that nanoparticle

toxicity is a function of some measurable physical characteristics and

that Structure Activity Relationship (SAR) might ultimately be a useful

tool for nanotoxicology. A panel of 18 nanoparticles was chosen to

include carbonbased materials and metal oxides. Comparative toxicity

of the nanoparticles in the panel was assessed, using a number of

different measures of apoptosis and necrosis; haemolytic effects and

the impact, on cell morphology were also assessed. Structural and

composition properties were measured including size distribution,

surface area, morphological parameters, metal concentration,

reactivity, free radical generation and zeta potential. Toxicity end

points were processed using principal component analysis and

clustering algorithms.

QSARToxicity(invitro)AnalysisforaPanelofEighteen NPs

42QualityNanoTrainingSchool

Nanooxicology 2012,47September,Beijing

Particles Qualifier Number Supplier Form

Carbonblack Printex 90 N1 Degussa DryDieselexhaustparticles EPA N2 EPA Dry

Nanotubes Japanese N3 VickiStone DryFullerene N4 Sigma Dry

PolystyrenelatexUnmodified N5 Polysciences Sus.

Amine N6 Sigma Sus.Carboxylated N7 Polysciences Sus.

Aluminiumoxide7nm N8 KrahnChemie Dry50nm N9 AHT* Dry300nm N10 AHT* Dry

Ceriumoxide N11 NAM** DryNickeloxide N12 NAM** DrySiliconoxide N13 NAM** DryZincoxide N14 NAM** Dry

TitaniumdioxideRutile N15 NAM** DryAnatase N16 NAM** Dry

SilverN17N18

NAM**NAM**

DrySus.*AHT:AlliedHighTech,**NAM:NanostructureandAmorphousMaterialInc

The Set ofEighteen NPs

43QualityNanoTrainingSchool

Nanotoxiclogy 2012,47September,Beijing

Particle shape was analysed using LEO 1530 Scanning Electron Microscope (SEM) or PhilipsCM20 Transmission Electron Microscope (TEM) aspect ratio and mean sizeSurface area and porosity were measured using TriStar 3000 BET. Thirteen measurements,including five surface areas based on different definitions, three pore volumes, three poresizes as well as the mean size and particle density.Particle size and size distribution were analysed using a Malvern MasterSizer 2000. Sizedistribution, and seven other size properties (mass diameter, uniformity, specific surfacearea, surface area mean diameter and three mass diameters)ThefreeradicalactivitiesweremeasuredbyEPR(Electronparamagneticresonance)using the spin traps, DMPO and Tempone H separately. Tempone H has partial selectivity fortrapping superoxide anions, whilst DMPO traps mainly hydroxyl radicalsParticle reactivity in solution, the dithiothreitol (DTT) consumption test as a reducing specieswas carried out. Since the DTT consumption test can only be conducted on dry powders, onlyfourteen of the panel of nanoparticles were assessed using this assay.

Metal Concentration water soluble concentration of ten heavy metals (Ti, V, Cr, Mn, Fe, Co,Ni, Cu, Zn and Cd) was measuredCharge:zpotentialwasmeasuredusingMalvernInstrumentsZetasizer Nano instrument

CharacterisationofPhysicochemicalProperties

44QualityNanoTrainingSchool

45

1313 BETmeasurements(for14drysamples)

22 SEM/TEMmeasurements (forallsamples)

77 laserdiffractionsizestatisticalmeasurements(MasterSizer 2000)forallsamples

8484 laserdiffractionsizedistribution measurements

22 EPRmeasurements(forallsamples)

11 reactivitymeasurement (fordrysamples)

1010 metalcontentmeasurements

TotalNumber ofMeasurements

Total:119Measurements

Particlename

BET Measurements M2000 Measurements

SEM Measurements Zeta measurements Other Measurements

N1 A(1,1),A(1,2),,A(1,12) B(1,1),B(1,2) C(1,1),C(1,2) D(1,1),D(1,2) E(1,1),E(1,2)

N2 A(2,1),A(2,2),,A(2,12 B(2,1),B(2,2) C(2,1),C(2,2) D(2,1),D(2,2) E(2,1),E(2,2)

N18 A(18,1),A(17,2), B(18,1),B(18,2) C(18,1),C(18,2) D(18,1),D(18,2) E(18,1),E(18,2)

QualityNanoTrainingSchool

Nanotoxicolgy 2012,47September,BeijingSEMandTEMimagesofthe18nanoparticlesamples

46QualityNanoTrainingSchool

Table2.Thewatersolubleconcentrationsofvariousmetalsforthe18nanoparticlesamplesNanotoxicology 2012,47September,Beijing

Particlenames MetalConcentration(g/g)Ti V Cr Mn Fe Co Ni Cu Zn Cd

CarbonBlackN1 0.000 0.057 0.000 0.305 0.000 0.002 0.263 1.161 9.687 0.008

DieselExhaustN2 2.320 0.312 16.47 20.81 208.4 0.508 4.940 2.235 3181 0.387

JapaneseNanotubesN3 0.000 0.100 0.000 0.161 13.10 0.000 0.299 0.123 0.460 0.001

FullereneN4 0.000 0.084 0.000 0.000 0.000 0.000 0.534 0.142 4.837 0.000

PolystyreneLatexBeadsN5 0.000 1.172 0.000 0.000 0.000 0.000 0.000 0.625 19.56 0.000

PolystyreneLatexBeadsN6 4.368 0.779 0.343 0.199 0.000 0.006 0.440 18.23 58.71 0.010

PolystyreneLatexBeadsN7 0.000 1.018 0.000 0.000 0.000 0.000 0.000 2.588 22.323 0.000

Aluminuim OxideN8 0.000 0.000 0.000 0.000 0.000 0.000 0.000 0.714 0.000 0.000

AluminuimOxideN9 0.000 0.031 0.000 0.000 0.000 0.000 0.056 0.099 4.987 0.000

AluminuimOxideN10 0.000 0.062 0.000 0.000 0.000 0.000 0.023 0.088 0.000 0.000

CeriumOxideN11 0.000 0.004 1.058 0.128 0.000 0.000 0.616 0.191 112.9 0.007

NickelOxideN12 0.000 0.056 0.000 0.000 0.000 0.042 16963 0.286 0.000 0.038

SiliconOxideN13 0.000 0.088 0.000 0.000 0.000 0.000 14.163 0.106 0.000 0.000

ZincOxideN14 0.000 0.096 0.000 0.000 0.000 0.000 1.336 0.047 8185 9.711

TitaniumDioxideRutile N15 0.000 0.083 0.000 0.000 0.000 0.000 0.286 0.276 39.623 0.000

TitaniumDioxideAnatase N16 0.000 0.104 0.000 0.160 0.000 0.006 0.692 0.303 402.3 0.000

SilverN17 0.000 0.111 0.000 0.503 0.000 0.000 0.475 0.597 17.867 0.019

SilverN18 0.000 0.096 0.000 0.256 0.000 0.059 0.561 162.8 194.8 0.037

Theheavy metalconcentration ofeach NP

47QualityNanoTrainingSchool

LEO1530GeminiFEGSEMScanningElectronMicroscopy

TheinformationaboutSize,Shape canbeobtainedfromSEMimages

CarbonBlackN1

Nanotoxicology 2012,47September,Beijing

MeansizeobtainedbymeasuringtheparticlesontheSEM/TEMimage:Mean_Size_SEMAspectratioobtainedfromSEM/TEMimages:Aspect_Ratio

CharacterisationofParticleSizeand Shape

48QualityNanoTrainingSchool

Nanotoxicology 2012,47September,Beijing

Micromeritics TriStar3000SurfaceAreaTechnique:GasAdsorptionMinimumSurfaceArea:0.01m2/g

SupportGases:NitrogenandHeliumMeasurementParameters:BETMeasurement

AdsorptionIsothermsDesorptionIsotherms

tplotsPoreVolumeAnalysisPoreSizeAnalysis

Surface Area (m2/g) Pore Volume (cm3/g) Pore Sizes ()

Single point surface area at P/Po =0.197 (m/g) SPSA

BET Surface Area: BET_SA Langmuir Surface Area: LM_SA BJH Adsorption cumulative surface

area of pores between 17.000 and3000.000 diameter :BJH_AC_SAOP

BJH Desorption cumulative surfacearea of pores between 17.000 and3000.000 diameter:BJH_DC_SAOP

Single point adsorption total pore volume of pores at P/Po = 0.98: SP_TP_VOP

BJH Adsorption cumulative volume of pores between 17.000 and 3000.000 diameter: BJH_AC_VOP

BJH Desorption cumulative volume of pores between 17.000 and 3000.000 diameter : BJH_DC_VOP

Adsorption average pore width: BET_PW

BJH Adsorption average pore diameter : BJH_A_PD

BJH Desorption average pore diameter: BJH_D_PD

Mean size calculated from BET surface area: Mean_Size_BET

Particle density: Density

**14drypowdersamples,noavailableforthreesuspensions(N5,N6,andN7)andthenewSilversample.

BETMeasurements for 14dry powder samples

49QualityNanoTrainingSchool

MalvernMastersizer 2000TypeofAnalyser:LaserDiffraction

SampleVolume:1000ml,150ml(HydroGandS)MeasurementRange:20nm 2000m

0123456789

10

0.01 0.1 1 10 100 1000 10000

Size (micron)

Prob

abili

ty (%

)

TheinformationaboutmeanSizeandSizedistribution,canbeobtainedfromthissizer.

Nanotoxicology 2012,47September,BeijingCharacterisationofNPsMastersizer

50QualityNanoTrainingSchool

D[4,3] theVolumeWeightedMeanorMassMomentMeanDiameter,representbyD[4,3]

Uniformity:ameasureoftheabsolutedeviationfromthemedian,representedbyUniformity

Specificsurfacearea:thetotalareaoftheparticlesdividedbythetotalweight,SSA_2000

D[3,2] theSurfaceWeightedMeanorSurfaceAreaMomentMeanDiameter,representedbyD[3,2]

D(v,0.5) thesizeinmicronsatwhich50%ofthesampleissmallerand50%islarger.ThisvalueisalsoknownastheMassMedianDiameterorthemedianofthevolumedistribution,representedbyD(0.5)

D(v,0.1) thesizeofparticlebelowwhich10%ofthesamplelies,representedbyD(0.1)

D(v,0.9) thesizeofparticlebelowwhich90%ofthesamplelies,representedbyD(0.9)

Sizedistribution:thereare100values.Amongthemthereare16valuesequaltozeroforallparticles.So84 sizedistributionattributesareusedforfurtheranalysis

Mastersizer 2000Measurements:91**

** Note:availableforallsamples

Nanotoxicology 2012,47September,BeijingCharacterisationofNPsMastersizer

51QualityNanoTrainingSchool

MalvernZetaSizerNanoinstrument Sizeranges:0.6nm 6m

ZetaPotentialSuitable:suspensions

Nanotoxicology 2012,47September,BeijingCharacterisationofNPsZeta Potential

52QualityNanoTrainingSchool

Cytotoxicity byLDHRelease

N6

N12

0.00

20.00

40.00

60.00

80.00

100.00

31.25 62.50 125 250

LDHRelease

Dose(g/ml)

N6

N14

0.00

5.00

10.00

15.00

20.00

25.00

31.25 62.50 125 250

Apop

tosis

Dose(g/ml)

Apoptosis

N6

N14

0.00

20.00

40.00

60.00

80.00

100.00

120.00

31.25 62.50 125 250

Viab

le

Dose(g/ml)

Viability

N3

N6

N14

0

10

20

30

40

50

60

70

80

31.25 62.50 125 250

Necrosis

Dose (g/ml)

Necrosis

0

10

20

30

40

50

60

70

80

90

100

Haemolysosatd

ose50

0g/ml

Proinflammation effects Haemolysis

53QualityNano Training School

Toxicity Results

54

LDH Apoptosis Viability

NecrosisHaemolysis

TotalToxicityMTT

QualityNanoTrainingSchool

Principal Component Analysis ofCytotoxicity Data

55

Figure: Plotofthefirstandsecond PCs basedonPCAanalysisoftoxicitydata

Japanesenanotubes

Zincoxide

Nickeloxide

Polystrene larex (Amine)

Themajorclustercontainslowtoxicityparticle

samples

Aminated beads(N6),zincoxide(N14),Japanese

Nanotubes (N3)andnickeloxide(N12)areseparate

fromthemajorcluster. Aminated beads(N6)hasthehighesttoxicityvaluesinnearlyallassayresults(LDH,apoptosis,necrosis,haemolytic,MTTandcellmorphologyassays).

Zincoxide(N14)hashightoxicityvalueinLDH,

apoptosis,necrosisandinflammationassays.

Japanesenanotubes (N3)havehightoxicityvalues

inviabilityandMTTassays.

Nickeloxide(N12)hasshownhighlytoxicinLDH

andhaemolyticassaysresults..

QualityNanoTrainingSchool

Principal Component Analysis ofCytotoxicity Data

Thedistinctionofthefourparticles,N6,N14,N3andN12,fromotherparticlesisclear.

56

NP6Average Toxicity Comparison NP14Average Toxicity Comparison

N14 has high toxicity value in LDH, apoptosis,necrosis and inflammation assays.

N12 hasshownhighlytoxicinLDHand haemolyticassaysresults

NP12Average Toxicity Comparison NP3Average Toxicity Comparison

N3 havehightoxicityvaluesinviabilityandMTTassays.

N6 hasthehighesttoxicityvaluesinnearlyallassayresults (LDH,apoptosis,necrosis, haemolytic,MTTandcellmorphologyassays).

QualityNanoTrainingSchoolThe toxicity results ofN6,N3,N12and N14which are likely to betoxic

57

The focus of the (Q)SAR Analysis is to identify the possible structural andcompositional properties which may contribute to the toxicity of zincoxide (N14), polystrene latex amine (N6), Japanese nanotubes (N3) andnickel oxide (N12).

Next Step:Principal Component Analysis ofDescriptors

PCA and clustering were then applied to the physicochemicaldescriptors with the hypothesis that the toxic nanoparticles might alsobe discriminated from the nontoxic particles based on the analysis ofthe measured physicochemical descriptors.

The theory is that there should be structural or/and compositionaldistinctions for these nanoparticles from the rest of the nontoxicparticles.

If this can be proved, further analysis can be conducted to identify thekey physicochemical descriptors that lead to the observed toxicity.

QualityNanoTrainingSchool

58

N2

N3

N12

N14

N6

N5

N7

Figure : PCA analysis of the physicochemicaldescriptors for the panel of 18 nanoparticles(excluding BET and DTT measured descriptorssince they are not available for the four wetsamples).

QualityNanoTrainingSchool

Principalcomponentanalysisofthestructuralandcompositionaldescriptorsforthe panelof18nanoparticles(excludingBETandDTTmeasureddescriptorssincetheyare notavailableforthefourwetsamples).

Principal Component Analysis ofDescriptors

The scaling operation was performed,3PCs

were then applied.

PC1PC2PC3threedimensionalplotshows

thatthe18nanoparticleswereclearlygrouped

intoclusters.

Thelargestclustercontainsparticlesthat

showedlowtoxicityvalues.

JapaneseNanotubes (N3),nickeloxide(N12)

andzincoxide(N14)anddieselexhaust

particles(N2)areoutsidethislowtoxicity

samplecluster.

There mustbeanotherreasonthatleadsto

hightoxicityvalueforN6

59

Toexaminetheclusteringresultsmoreclearly,PC1PC2, PC1PC3andPC2PC3twodimensionalplotsarealsoexamined .ItisPC2thatseparates N3(Japanesenanotubes),N12(nickeloxide)andN14(zincoxide) fromtherestsamples,ascanbeseenfromPC1PC2andPC2PC3plots.

QualityNanoTrainingSchool

Results

60

Figure: PC1,PC2andPC3ContributionplotsbasedonPCAanalysisofthestructuralandcompositionaldescriptorsforthepanelof18nanoparticles

QualityNanoTrainingSchool

Results

The contribution plot for, PC1, PC2 andPC3 is shown in the Figure, we are mainly

interested in the contribution to PC2 (solid

red colour), since it is PC2 that

discriminated Japanese nanotubes (N3)

nickel oxide (N12) and zinc oxide (N14)

from the rest particle samples.

The largest contributions to PC2 are: aspectratio measured by SEM imaging, and volume

weighted mean, uniformity, and D(0.9), all

measured by Mastersizer, as well as Ni and

Zn metal contents

Further analysis is required.

61

N2

N 3

N 12

N14

Inordertodetermineexactlywhichphysicochemicaldescriptors leadtohightoxicityvaluesofN3,N12,N14and N6,several PCanalysis wereperformed.

Forthe14dryparticles,PCanalysiswithBETandDTTmeasurements gavesimilarresultwith the first PCanalysisinwhich BETand DTTdatawasexcluded.

QualityNanoTrainingSchool

Figure: Principalcomponentanalysisofthestructuralandcompositionaldescriptors(allsuchdescriptorsincludingBETandDTTmeasurements)forthe14drynanoparticles.

Results

62

Figure: PC1andPC2contributionplotsforPCAanalysisof thecompositionaldescriptors

NicontentmadethemostimportantcontributiontoPC1,suggestingthatnickelcontentmaybethemainreasonforthehightoxicityof N12(nickeloxide). While Zncontentmadethemostobvious contributiontoPC2, suggesting thatzinccontentmaybethereasonfor the high toxicity of N2(dieselexhaustparticles)andN14(zincoxide).

QualityNanoTrainingSchool

Results

To identify the relativecontribution of each metalconcentration to PC1 andPC2, the contribution plot toPC1 and PC2 is shown in theFigure.

Table2.Thewatersolubleconcentrationsofvariousmetalsforthe18nanoparticlesamplesNanotoxicology 2012,47September,Beijing

Particlenames MetalConcentration(g/g)Ti V Cr Mn Fe Co Ni Cu Zn Cd

CarbonBlackN1 0.000 0.057 0.000 0.305 0.000 0.002 0.263 1.161 9.687 0.008

DieselExhaustN2 2.320 0.312 16.47 20.81 208.4 0.508 4.940 2.235 3181 0.387

JapaneseNanotubesN3 0.000 0.100 0.000 0.161 13.10 0.000 0.299 0.123 0.460 0.001

FullereneN4 0.000 0.084 0.000 0.000 0.000 0.000 0.534 0.142 4.837 0.000

PolystyreneLatexBeadsN5 0.000 1.172 0.000 0.000 0.000 0.000 0.000 0.625 19.56 0.000

PolystyreneLatexBeadsN6 4.368 0.779 0.343 0.199 0.000 0.006 0.440 18.23 58.71 0.010

PolystyreneLatexBeadsN7 0.000 1.018 0.000 0.000 0.000 0.000 0.000 2.588 22.323 0.000

Aluminuim OxideN8 0.000 0.000 0.000 0.000 0.000 0.000 0.000 0.714 0.000 0.000

AluminuimOxideN9 0.000 0.031 0.000 0.000 0.000 0.000 0.056 0.099 4.987 0.000

AluminuimOxideN10 0.000 0.062 0.000 0.000 0.000 0.000 0.023 0.088 0.000 0.000

CeriumOxideN11 0.000 0.004 1.058 0.128 0.000 0.000 0.616 0.191 112.9 0.007

NickelOxideN12 0.000 0.056 0.000 0.000 0.000 0.042 16963 0.286 0.000 0.038

SiliconOxideN13 0.000 0.088 0.000 0.000 0.000 0.000 14.163 0.106 0.000 0.000

ZincOxideN14 0.000 0.096 0.000 0.000 0.000 0.000 1.336 0.047 8185 9.711

TitaniumDioxideRutile N15 0.000 0.083 0.000 0.000 0.000 0.000 0.286 0.276 39.623 0.000

TitaniumDioxideAnatase N16 0.000 0.104 0.000 0.160 0.000 0.006 0.692 0.303 402.3 0.000

SilverN17 0.000 0.111 0.000 0.503 0.000 0.000 0.475 0.597 17.867 0.019

SilverN18 0.000 0.096 0.000 0.256 0.000 0.059 0.561 162.8 194.8 0.037

Theheavy metalconcentration ofeach NP

63QualityNanoTrainingSchool

64

IdentifyingthefactorthatdiscriminatesN6(polystrene latexamine)fromN5(unmodified)andN7(carboxylated)

PCAanalysis:PCAanalysisofmetalcontentsonly,orPCAanalysisofstructuraldescriptorsonly,orPCAanalysisofbothstructuralandcompositionaldescriptorstogether,couldnotdiscriminateN5,N6andN7fromthelargestclustersofnontoxicparticlesamples.

PCAConclusion:neitherstructuraldescriptorsnormetalcontentsmadeN6differentfroN5andN7intoxicity

ItwasatthisstageofanalysisthatwesuspectedthatthereareotherpossiblereasonsfortoxicityspecifictoN6:eitheritwasduetothechargedifference,orduetoN6beingamine,orboth.

zetapotentialforallthreepolystyrenelatexsamples.N6(amine) 37.8,37.5,and40.3(mV), positiveN5(unmodified) 36.2,38.8and36.8, negativeN7(carboxylated) 54.9,55.3,and58.6., negative

AccordingtoVerma etal.anyparticlewithpositivechargeislikelytointeractelectrostatically sincemostcellsurfacesandotherbiologicalmembranesarenegativelychargedunderphysiologicalconditions(Verma etal.2008).Therefore,itismostlikelythatthelargepositivechargeofN6contributeditsobservedhightoxicity,despiteitstructurallysimilartoN5andN7.

Conclusions

QualityNanoTrainingSchool

65

TheSARanalysisgeneratesthefollowingconclusions:1. Themostlikelyreason ofhightoxicityforJapanesenanotube(N3)isitsshape(high

aspectratio).2. Themostlikelycauseofhightoxicityforzincoxide(N14)andnickeloxide(N12)is

their highcontentsofzincandnickel.3. Although,Aminated beads(N6)hasthehighesttoxicityvaluesinnearlyallassay

results,itwas always groupedcloselywiththeothertwopolystyrenelatexbeads,unmodifiedbeads(N5)andcarboxylated beads(N7),whichshowednotoxicity.Therefore,thezetapotentialofthethreepolystrene latexwas measured.TheresultsuggeststhatthelargepositivechargeofN6contributed to itsobservedhightoxicity.

TheworkhasshownthatSARbasedonmolecular descriptors isausefultoolfordescribingfactorsinfluencingthetoxicityofnanoparticles.

Apanelofeighteennanoparticlesisstillconsideredtobetoosmall,therefore,aQSARmodelthatcanbeappliedtopredictthetoxicityofnewnanoparticles cannot bedeveloped with this small setofdata.

QualityNanoTrainingSchool

Conclusions