ContentslistsavailableatScienceDirect
Journal
of
Pharmaceutical
and
Biomedical
Analysis
jo u r n al ho me p a g e :w w w . e l s e v i e r . c o m / l o c a t e / j p b a
Simultaneous
automated
image
analysis
and
Raman
spectroscopy
of
powders
at
an
individual
particle
level
Andrea
Sekulovic
a,b,
Ruud
Verrijk
b,
Thomas
Rades
a,
Adam
Grabarek
c,d,
Wim
Jiskoot
c,d,
Andrea
Hawe
c,
Jukka
Rantanen
a,∗aUniversityofCopenhagen,DepartmentofPharmacy,Denmark bDrReddy’sResearch&DevelopmentB.V.,Leiden,TheNetherlands cCoriolisPharma,Martinsried,Germany
dLeidenUniversity,DivisionofBioTherapeutics,TheNetherlands
a
r
t
i
c
l
e
i
n
f
o
Articlehistory: Received25August2020
Receivedinrevisedform26October2020 Accepted27October2020
Availableonline4November2020 Keywords:
Polymorphism Ramanspectroscopy Imageanalysis
Multivariatedataanalysis
Partialleastsquares-discriminantanalysis Modelling
a
b
s
t
r
a
c
t
Solidformdiversityofrawmaterialscanbecriticalfortheperformanceofthefinaldrugproduct.In thisstudy,Ramanspectroscopy,imageanalysisandcombinedRamanandimageanalysiswereutilized tocharacterizethesolidformcompositionofaparticulaterawmaterial.Ramanspectroscopyprovides chemicalinformationandiscomplementarytothephysicalinformationprovidedbyimageanalysis.To demonstratethisapproach,binarymixturesoftwosolidformsofcarbamazepinewithadistinctshape,an anhydrate(prismshaped)andadihydrate(needleshaped),werecharacterizedatanindividualparticle level.Partialleastsquaresdiscriminantanalysisclassificationmodelsweredevelopedandtestedwith known,gravimetricallymixedtestsamples,followedbyanalysisofunknown,commerciallysupplied carbamazepinerawmaterialsamples.Classificationofseveralthousandsofparticleswasperformed, anditwasobservedthatwiththeknownbinarymixtures,theminimumnumberofparticlesneededfor thecombinedRamanspectroscopy–imageanalysisclassificationmodelwasapproximately100particles persolidform.Thecarbamazepineanhydrateanddihydrateparticlesweredetectedandclassifiedwith aclassificationerrorof1%usingthecombinedmodel.Further,thisapproachallowedtheidentification ofrawmaterialsolidformimpurityinunknownrawmaterialsamples.Simultaneousautomatedimage analysisandRamanspectroscopyofpowdersatanindividualparticlelevelhasitspotentialinaccurate detectionoflowamountsofunwantedsolidformsinparticulaterawmaterialsamples.
©2020PublishedbyElsevierB.V.
1. Introduction
Polymorphismofdrugcompoundsinvolvestheoccurrenceof differenttypesofpackingofthesamemoleculeinacrystal lat-tice [1]. The expression “solidform” can be used as a broader termtodescribenotonlycrystallinesinglecomponentsystems, butalsoamorphousmatterandbinarysystems,suchassalts, sol-vates,cocrystallineandcoamorphoussystems[2].Differentsolid formsofadrugcompoundcanbecriticalforthehealthoutcomeof apatientbecausetheymayaffectproductperformance,especially whenexhibitingadifferentparticlesize,shape,solubility, dissolu-tionrateandbioavailability[3,4].Thesecriticalmaterialattributes canthusaffectproductquality,safetyandefficacy[5].Unexpected solid formchanges, suchasmetastable polymorphs[6],elusive
∗ Correspondingauthor.
E-mailaddress:jukka.rantanen@sund.ku.dk(J.Rantanen).
crystalforms[7] andunintentionalseeding causedbyvery low amountsofanunwantedpolymorph[8],havehadnegativeeffects ontheavailabilityofotherwiseaffordabledrugs.Anexampleofa drugproductwithadetrimentaluncontrolledsolidformchange wasNorvir® (ritonavir),leavingAIDSpatientstemporarily
with-outatreatment[8].Toensurethatthepolymorphismlandscape isproperly exploredby theindustry,the regulatoryauthorities haveissuedguidancedocumentsonsolidformcharacterization andcontrol[9].Forreasonssuchastheonesmentionedabove,the solidformdiversityofparticulatematter,bothinsolid,semi-solid andliquidproducts,isofparticularinteresttothe pharmaceuti-calindustrywhenaimingforamoredetailedproductandprocess understanding.Solidformscreeninghasbecomeanindustrial prac-ticetocopewiththesechallengesandin thiscontext,different methodsforgeneratingthemaximalnumberofnewformsaswell ashighthroughputanalyticalmethodsforquantificationand detec-tionhavebeendeveloped[10].
https://doi.org/10.1016/j.jpba.2020.113744
tiatedintensiveanalyticsinsearchingforverylowamountsofa givensolidform[8].
Conventionalapproachestodetectandquantifylowlevelsof solid formimpurities arebased ontheanalysisofbulk materi-als[12].Herewereportananalyticaltechniqueatanindividual particlelevel for thedetection and quantificationof solid form diversity.Crystalstendtogrowinaspecificcrystallographic direc-tion[13]anddifferentpolymorphsandcrystallization/processing conditionstypicallyresultindifferentcrystalmorphologies[14]. Quantitativeassessmentofparticlemorphologyhasevolvedwith technological advancements in instrumentationand computing power.Methodsbasedonautomatedparticletracking, bright/dark-fieldimagingandimageanalysis(IA)arecapableofparticlesizeand shapeanalysisofevenhundredsofthousandsofparticleswithina reasonabletimeframe[15].Particlemorphologycouldthereforebe usedforsolidformassessment,potentiallyresultinginafastand sensitiveanalyticaltechnique.
Ithasbeenestimatedthatabout90%ofstudiesrelatedto poly-morphismuseatleasttwosolid-stateanalyticaltechniques[16]. Theuseofacombinationofcharacterizationtechniquesenables scientificinsightintothecomplexityofthesephenomenawitha higheraccuracy[17].OurstudyevaluatesRamanspectroscopy,IA andthecombinationofRamanspectroscopyandIAatasingle par-ticlelevelinordertodetectandquantifyalowamountofasolid formimpurityofacrystallinedrugmaterial.Partialleast squares-discriminantanalysis(PLS-DA)isawell-establisheddataanalytical methodthatcombinesdimensionalityreductionandhigh predic-tioncapability.PLS-DAiscommonlyusedforvariableselectionas wellaspredictiveanddescriptiveclassificationmodeling.The PLS-DAalgorithmisapplicableforanalyzinghighdimensionaldataand doesnotassumethedatatofitanydistribution,makingitsuitable forimbalanceddatawithahighnumberofvariables(n>1000)[18]. Byusingcrystallinecarbamazepine(CBZ)anhydrate(AH)and dihy-drate(DH)asmodelsolidforms,weaimtoquantifysolidformsin thesebinarymixtures,ultimatelyevenata singleparticlelevel. QuantificationbasedonIA,Ramanspectroscopyoracombination thereofiscomparedandastrategyisproposedfordetectionoflow amountsofanunwantedsolidform.
2. Experimentalsection 2.1. Materials
Carbamazepine(CAS298−46-4)waspurchasedfromthree differentcommercialsources:TokyoChemicalIndustry,Co.,LTD (Tokyo, Japan), Hawkins, Inc. (Minneapolis, MN,USA) and Car-bosynth(Berkshire,UK).Methanol99.8%(67−65-1)waspurchased fromSigmaAldrichCo.(St.Louis,Missouri,MO,USA).All chemi-calsusedwereofanalyticalreagentgradeorhigher.Highlypurified water (Milli-Q, MilliporeInc., Denver,Massachusetts, USA) was usedinallofthestudies.HydrophilicPTFEfilters,Omnipore(with a0.45mporesizeand47mmdiameter),werepurchasedfrom Merck(Darmstadt,Germany).
imizeadsorptiontoglassvials.Threedifferentphysicalmixturesof CBZDHandCBZAHwereprepared:testsample1,composedof20 %DH(w/w)(4mgCBZDHmixedwith16mgCBZAH),testsample 2,composedof50%DH(w/w)(10mgCBZDHmixedwith10mg CBZAH)andtestsample3,composedof80%DH(w/w)(16mgCBZ DHmixedwith4mgCBZAH).
2.2.2. Characterizationofthecommerciallysupplied carbamazepinerawmaterialsamples
CommerciallysuppliedCBZrawmaterialsampleswere char-acterizedasindicatedinSupportingInformation,Section1Cand 8.
2.2.3. RamanspectroscopyandIA
Pure recrystallized, test set, and commercially supplied CBZ samples were characterized by using an automatic optical microscope with an integrated Raman spectrometer (Malvern Panalytical,Worcestershire,UK). Thesetupallows forthe mea-surementof Raman spectraand IA at a single particlelevel of uptoseveralhundredthousandparticleswithinatimeframeof 1–10hours. Theinstrument details aregiven in theSupporting InformationSection1D.
Each sample was dispersed onto a glass plate by using an automated sample dispersion unit (SDU) and scanned for the particle size and shape IA of about 100,000 particles with a microscope objective that covers the particle size range of 1.8m–100m.Theparticlesizefilter,equivalentcircular diame-ter(ECD)7m–300m,wasusedforRamanspectroscopyanalysis ofatleast1000particles(TableS1,SupportingInformation).A par-ticlesizefilterwasappliedtoensuresufficientRamanintensity andadequateresolutionforparticleshapeanalysisandtoavoid aggregates.Alowfiltercutoffof>7mwasusedtoensure suffi-cientRamanintensityandadequateresolutionforparticleshape analysis.ThespotdiameteroftheRamanlaserbeamwas approxi-mately3m.Theobtainedspectrahadaresolutionof6cm−1and comprisedofaRamanwavenumberrangeof150−1850cm−1.
2.2.4. Preparationofdataformodeling
Theraw data wasexported ascsv files, data matrices were preparedbyusingR(Rstatisticalsoftwareenvironment version 3.4.3)andalldatawasimportedintoMATLAB(MATLABVersion 8.6.0.267246R2015b) for subsequent multivariatemodeling by usingthePLSToolbox(PLSToolboxVersion8.1.120,131).
Foreachrecrystallizedpuresolidform,atleast1000particles wereanalyzedbyRamanspectroscopyandIA(TableS1,Supporting Information).Subsequently,1241particlesofeachsolidformwere mergedinsilicointoonedatamatrixwithatotalof2482particles representing50%(n/n)CBZAHand50%(n/n)CBZDH,andused formodeling(Fig.1).
Ramanspectroscopyrawdatawascomposedof1701 wavenum-bersandIArawdataof12particlesizeandshapedescriptors,per particle(Fig.1).Both,RamanspectroscopyandIAdatawas ran-domlypartitionedinto2/3and1/3ofthetotalnumberofparticles formodeltrainingandvalidation.Ramanspectroscopydatawas
Fig.1. Partialleastsquares-discriminantanalysis(PLS-DA)trainingandvalidationdatasetscomposedofIAandRamanspectroscopydataofrecrystallizedpureCBZanhydrate (AH)anddihydrate(DH)samples.
smoothedbyusingastandardnormalvariate(SNV)algorithmand meancentered;IAdatawasauto-scaledandmeancentered(Fig.1). 2.2.5. Ramanmodel
AsupervisedmultivariateRamanclassificationmodelwas cre-atedwiththepartialleastsquares-discriminantanalysis(PLS-DA) algorithm. The number of latentvariables (LVs) of the Raman PLS-DAmodel wasdeterminedbyusinginternalvalidation, e.g. calibration/training and10-foldvenetianblindscrossvalidation. Raman PLS-DAmodel quality wasexpressedbythe sensitivity, specificity, accuracy, precision and classification error. Perfor-mance of themodel was evaluated with thetest samplesand commerciallysuppliedCBZrawmaterialsamples.DifferentRaman PLS-DAmodelswerecreatedwithastepwiseincreasingnumberof randomlysampledpurerecrystallizedparticlesofbothsolidstate formsofthedrug.Todeterminehowmanyparticleswereneeded foraclassificationmodel,theiterativeprocessofrandomsampling continuedwithanincreasingnumberofparticles,untilthe pre-dictednumberofhydrateparticlesvisuallyleveledofftoaconstant value.
2.2.6. IAmodel
TheIAPLS-DAmodelwascreatedthesamewayastheRaman modelbyusingthesameparticles.Variableimportancein projec-tion(VIP) scoreswerecalculatedwiththePLStoolboxsoftware [19].TheminimumnumberofparticlesneededfortheIAPLS-DA modelwasassessedinthesamemannerasfortheRamanPLS-DA model.
2.2.7. Combinedmodel
ThecombinedPLS-DAmodelwascreatedwithconcatenated, auto-scaledandmeancenteredRamanPLS-DAscorevaluesforthe initialLVsandIAparticlesizeandshapeparmeters.Variable impor-tance,modelingandassessmentofthenumberofparticleswere evaluatedthesamewayasfortheIAmodel.
3. Resultsanddiscussion 3.1. Modeldevelopment 3.1.1. Ramanmodel
The Raman PLS-DA model wasoptimal withthree LVsthat cumulativelyused76.9%ofthevariationinthedatatoclassify CBZAHandCBZDH(TableS2,SupportingInformation).Separation ofCBZAHandCBZDHclassesisvisualizedwiththescoresplot ofthethreeLVs(FigureS4,SupportingInformation).Theloadings plot(FigureS5,SupportingInformation)indicatesthattheRaman wavenumberrangebetween1400cm−1and1600cm−1isthemost significantforclassification,inagreementwiththeliterature[20]. TheRamanmodelclassifiedCBZAHsolidformwithatrue pos-itiverate(TPR,sensitivity)of98.7%andatruenegativerate(TNR,
specificity)of97.6%(TableS3,SupportingInformation).CBZDH classificationsensitivity and specificity were97.6%and 98.7% respectively.TheaccuracyoftheRamanmodelwasonaverage98.2 %,resultinginaclassificationerrorof1.8%(TableS3,Supporting Information).TheCBZDHclassificationprecisionwas98.7%, com-paredto97.7%fortheCBZAHclassificationprecision.Intotal,the predictedCBZDHclasswassmallerandhadlessfalsepositives, resultinginahigherpercentageoftruepositivescomparedtoCBZ AH.
The lower CBZ DH sensitivity compared to CBZ AH could havebeencausedbythespherical3mdiameterlaserspotnot optimallyinteractingwiththeneedle-shaped CBZDH particles, oftenrecordingthebackgroundscatteringalongwiththe parti-cle,resultinginalowsignal-to-noiseratio(FigureS3,Supporting Information).CBZDH needleparticleshadaminimumwidthof 1.2m,whichwassmallerthantheminimumwidthof2.7mof theprismshapedCBZAHparticles.Thedifferenceintheminimum widthcouldhavebeenimportantforthedetectionofthe parti-clesbythecameraandthelaser,and particlescouldhavebeen missedwhenthemagnificationwaschangedafterparticle track-ingbutbeforetheRamanspectroscopyanalysis,observedat20x and50x,respectively.AhigherCBZDHspecificitycomparedtoCBZ AHiscausedbyalownumberofclassifiedCBZAHparticlesactually beingCBZDH(TableS4,SupportingInformation).CBZAHparticles werenotmisclassifiedbecausetheirRamanspectradidnothavea highbackgroundnoise.
Manyspectroscopic methods canbeconsidered forfast and sensitive approaches to solid form detection. Terahertz pulsed spectroscopyhasbeenreportedtodetectsolidformswithalimitof detectionofapproximately1%[21].Similarly,solid-statenuclear magneticresonance(ss-NMR)hasadetectionlimitof1%[22].The approachpresentedhereisbasedondetectingsolidform impu-ritiesatasingleparticlelevel,whichisafundamentallydifferent approachtothesebulkmethods.Whenanalyzingahighnumberof particles,thelevelofdetectioncanbecomeaslowasthe measure-menttimeanddatastorageallows,whichultimatelymayevenlead tofindingasingleimpurityparticleamongstmillionsofparticles [23].
3.1.2. Imageanalysis(IA)model
TheIAbasedPLS-DAmodelhadoptimalclassificationwithfour LVsthatcumulativelyused89.5%ofthevariationinthedata(Table S5,SupportingInformation).LessclearvisualseparationofCBZAH andCBZDHclasseswasachievedbasedonIAmodel compared toRamanmodel (FigureS6,Supporting Information).The vari-ableimportanceprojection(VIP)plotindicatedtheintensitymean (IM),aspectratio(AR)andHSasthemostsignificantparametersin explainingtheIAmodel(FigureS7,SupportingInformation).CBZ AHparticlesappeareddarkercomparedtoCBZDHparticles(Figure S2,SupportingInformation),resultinginalowerIM.IMisthemean greyscalevalueofthepixels(FigureS8,SupportingInformation)
Fig.2. CBZAH(red)andCBZDH(blue)classificationachievedbythecombinedRaman-IApartialleastsquarediscriminantanalysis(PLS-DA)model:Scorevaluesforthe first,secondandthirdlatentvariable(LV)oftheauto-scaledandmeancenteredtrainingdataset,n=1654(a)andthevalidationdatasetn=828(b).Theellipsoidsindicate the95%confidenceintervals.
rangingfrom0,representingblack,to255,representingwhite.This IMparameterisrelatedtotheopticalpropertiesofthecrystals, ulti-matelydeterminedbythecrystalpackinginthesemodelsystems. CBZDHparticleswereneedles,i.e.,hadaremarkablylargerlength thanwidthcomparedtoCBZAHparticles(FigureS2,Supporting Information),resultinginahigherARfortheCBZAHparticles (Fig-ureS8,SupportingInformation).ARistheratiobetweenthewidth andthelengthoftheparticles(EquationS6,Supporting Informa-tion).BecauseofthehigherAR,CBZDHparticleshadalowerHS circularitycompared toCBZAH particles(FigureS8,Supporting Information).HScircularityisdefinedashighsensitivity circular-ity,orsquaredcircularity,wherecircularityisthemeasureofhow closeaparticleistoaperfectcirclebasedontheratiobetweenthe areaandtheperimeter(EquationS8,SupportingInformation).
TheIAmodelclassifiedCBZAHwithasensitivityof98%anda specificityof93%.CBZDHwasclassifiedwithasensitivityof93 %andaspecificityof98%(TableS3,SupportingInformation).CBZ AHandCBZDHclassificationwaslesssensitivewhenbasedonIA comparedtoRamanspectroscopy.Alsospecificitywaslowerand moreCBZAHandCBZDHparticlesweremisclassifiedbasedonIA comparedtoRamanspectroscopy.Moreover,theaverageIAmodel accuracyof95.4%waslowercomparedtotheRamanmodel accu-racy,resultinginaclassificationerroroflessthan5%.TheIAmodel precisionofCBZAHclassificationwas93.1%,and97.9%forCBZDH classification.Forboth,theIAandRamanmodel,precisionofCBZ AHclassificationwaslowercomparedtoCBZDHclassification.The numberofmisclassifiedCBZDHparticleswashighercomparedto CBZAH.
EventhoughtheIM,ARandHScircularityarenotsignificantly differentbetweenCBZAHandCBZDHparticles(FigureS8, Sup-portingInformation),thecombinationofthemcanbeusedbythe multivariatePLS-DAmodeltodiscriminatebetweentheCBZAH andCBZDHclasses.PLS-DAisespeciallyapplicabletomodelingof correlatedvariablessuchasparticleshape,i.e.theARandHS cir-cularity.BoththeIAandtheRamanmodelclassifyCBZDHwitha lowersensitivityandhigherspecificitycomparedtoCBZAH.This maybebecausenotallCBZDHparticleswereneedleshapedand notallCBZDHparticlesthatwereneedleshapedwereobserved asneedlesdueto2Dimaginginnot-longitudinaldirection.Alsoif partsofCBZDHparticleswereoutoffocus,theycouldhaveskewed theoverallmorphologicalparametersoftheseparticles.IAresulted inabetterdetectionofCBZAHparticlescomparedtoCBZDHdueto thespecificshapeofCBZAHparticles,whichmakesitpossiblefor thecameratofocus/detecttheprismshapedparticlesmoreeasily. TheIAmodelresultedinlessmisclassifiedCBZAHparticles com-paredtoCBZDHparticles(TableS6,SupportingInformation).This maybebecausenotmanyCBZAHparticleswereneedleshapedor canbemistakenlyidentifiedassuchbytheIA.
Classificationerroroflessthan5%canbeattributedtotheclear visualdifferencebetweenCBZAHandCBZDHparticlesandthe
informationdensityofferedbyautomaticIA.Generallyusedimage basedparticlecharacterizationmethodsrelatedtothequalityofthe APIareopticalmicroscopyandscanningelectronmicroscopy.Both methodshaveevolvedwiththecurrentimagingsystemstowards automatedtimelyanalysisofthousandsofparticlesleadingtoa largeramountofdataonparticulatesystems[24].Adisadvantage ofusingtheIAmodelwasmisclassificationofCBZDHparticlesthat arenotneedleshaped,whileadisadvantageoftheRamanmodel wasthedifficultyoffocusingthelaserontheneedleshapedCBZ DHparticles.Therefore,acombinationofRamanspectroscopyand IAispotentiallyabletoclassifyCBZDHwithalowerclassification errorcomparedtoRamanmodelorIAmodelalone,makingthese analyticaltechniquescomplementary.
3.1.3. Combinedmodel
AcombinedmodelwascreatedwithfourLVsthatcumulatively use77.0%ofthevariationinthedataforclassification(TableS7, SupportingInformation).Theaimofusingthecombinedmodelwas toenableimprovedclassificationbyincreasedinformationdensity comparedtotheRamanorIAmodelalone.Theclassificationinto CBZAHandCBZDHgroupsbasedonthecombinedmodelisshown inFig.2.CBZAHandCBZDHclassesarevisuallynarrowercompared totheRamanmodel(FigureS4,SupportingInformation)andmore clearlyseparatedcomparedtotheIAmodel(FigureS6,Supporting information).
BasedontheVIPanalysis,Ramanspectroscopyisthemost sig-nificantvariableof thecombinedmodel (FigureS9,Supporting Information).ThecombinedmodelandRamanmodelarebasedon 77%explainedvarianceandtheIAmodelisbasedon90%explained varianceindicatingtheimportanceofthequalityofthe informa-tion,notthequantity.Theamountoftheexplainedvarianceisnot relatedtotheLV’sabilitytodiscriminatebetweentheclasses[25]. Ramanspectroscopyprovideschemicalinformationcomparedto thephysicalinformationprovidedbyIA.Consequently,the com-binedmodelhashighermodelqualityparameterscomparedtothe RamanmodelandtheIAmodelalone(Fig.3),demonstratingthe complementarynatureofthesetwoanalyticaltechniques. 3.2. Performanceofthemodelsusingtestsamples
TheperformanceoftheIAmodel,theRamanmodelandthe combinedmodelwastestedwiththephysicallymixedtest sam-ples.TheincreaseintheCBZDHcontentbetweenthetestsamples, i.e.20%to50%to80%(w/w)CBZDH,wasobservedwiththeRaman modelas19.3%to48.3%to75.0%(n/n)CBZDH,as17.1%to30.2% to63.3%(n/n)CBZDHwiththeIAmodel,andas18.7%to46.9%to 79.2%(n/n)CBZDHwiththecombinedmodel(Table1,FiguresS10 -S11,SupportingInformation).Fromtheaboveitisclearthatthe IAmodelclassifiedalowerCBZDHcontentforallthetestsamples
Fig.3. ModelqualityattributesoftheIA,Ramanandcombinedpartialleastsquares discriminantanalysis(PLS-DA)models.
comparedtotheRamanmodel,whereastheRamanmodelandthe combinedmodeldeterminedacomparableCBZDHcontent.
Bycombiningchemicalandphysicalinformationinthe com-bined Raman- IA model,thePLS-DA approach gainsimproved predictiveability.Ontheonehand,asstatedabove,achallenge wasobservedwithusingRamanspectroscopywhenthespotsize ofthelaserbeaminthecurrentsetupwasnotoptimalfortheneedle shapedCBZDHparticles.Ontheotherhand,IAhadchallengeswith thedetectionofCBZDHparticlesthatwerenotneedleshaped.The combinedmodelwasabletohandleCBZDHparticlessuccessfully becausethesemethodsarecomplementary.NeedleshapedCBZDH particlesweresuccessfullyclassifiedbyIAandnot-needleshaped CBZDHparticlesbyRamanspectroscopy.Havingtheclassification errorbroughtdownto1%withthecombinedmodel,assessing thousandsofparticleswilltheoreticallyresultinthedetectionof verylowlevelsofCBZAHorCBZDHwithanaccuracyof99%(Table S3,SupportingInformation).
Furthermore,aminordifferencebetweentheweightbasedCBZ DHcontentandthenumberbasedCBZDHcontentwasexpected, astheweightofthetestsamplesmaycontaina different num-berofparticlesdependingontheirrespectivedensityandvolume. ThedensityofCBZDHandAHis1.29g/cm3[26]and1.34g/cm3
[27],respectively.Thevolumeoftheparticlesneedstobeestimated basedonassumptionsofthethirddimensionoftheparticlesas opticalmicroscopyrecordsonlytwodimensionalimages.When assumingthesphericalequivalentvolume,numberbased percent-ages17.1%,19.3%and18.7%DH(n/n)resultedintherespective weightbasedpercentages28.7%,32.4%and31.4%DH(w/w).When assumingthethirddimensionwasequaltothewidthofthe par-ticles,theweightbasedpercentageswere24.3%,27.4%and26.6 %DH(w/w).Thesecalculationexampleshighlightthechallengeof transformingnumbertoweightbasedpercentages;careful consid-erationisneededwhencomparingnumericvaluesofparticlesize resultsbasedondifferentanalyticalmethods.
3.3. Performanceofthemodelsusingcommerciallysupplied (unknown)samples
TheIA,Ramanandcombinedmodelsweretestedforreallife performancebyassessingtheDHcontentofcommerciallysupplied CBZrawmaterialsamples.Thepurposewasnottoquantifythe absoluteDHcontentinthesesamplespersebuttodevelopa
mul-Fig.4. PurerecrystallizedCBZAH(red),CBZDH(blue),threedifferentcommercially suppliedcarbamazepine(CBZ)rawmaterialsamples(SupplierI,IIandIII) classifi-cationachievedbytheprincipalcomponentanalysis(PCA)model:Scorevaluesfor thefirst,secondandthirdprincipalcomponent(PC)oftheSNVcorrectedandmean centeredbulkRamanspectroscopydata.Theellipsoidsindicatethe95%confidence intervals.
tivariatebasedmodelingstrategyforsuchmeasurementsinthe futurebycomparingtheperformanceandrobustnessofthe mod-els.Inordertooptimizethetimeusedforthesemeasurements,the particlesizefiltercutoffusedfortheanalysisofthecommercially suppliedCBZ rawmaterial sampleswas40m.Withthis filter cutoffthemaximumECDoftheanalyzedparticleswas39.99m, whichmeansthattheappliedfilterinfluencedtheabsolute compo-sitionofthesamplesdeterminedbythemodels.Forfutureresearch ontheabsolutecompositionofthesamples,werecommendnotto useafiltercut-offtoavoiditspossibleinfluenceontheabsolute compositionofthesamples.
Commerciallysupplied CBZ sampleshad an inhomogeneous appearancewhenobservedbyscanningelectronmicroscopy(Fig.5 d,e, f;SupportingInformationSection1C).Thisis typicallyan indicationofvariationinthesolidformcomposition,andinthis case,thesesamplescontainedamixtureofparticlehabitssimilar tobothCBZAHandDH(Fig.5g,h).Potentialsolidformdiversityof commerciallysuppliedCBZmaterialhasbeenalsoreportedearlier [3].
SolidformdiversityofthecommerciallysuppliedCBZsamples wasinitiallyconfirmedwithunsupervisedmultivariateanalysisof thebulkRamanspectrawithprinciplecomponentanalysis(PCA) (Fig.4).
InFig.5a,b,c,this solidformvariationis visualizedbythe combinedRaman-IApartialleastsquarediscriminantanalysis (PLS-DA)model asa scoreplotofthefirstthreeLVs.Inhomogeneity ofthesesamplesisclearlyindicatedbythemixtureofthepure CBZ-AHorCBZDHsolidforms.
TheRaman,IAand combinedmodelsweretestedfor perfor-mancebyassessingtheCBZDHcontentofcommerciallysupplied CBZrawmaterialsamples.BasedontheRamanmodel,theIAmodel andthecombinationmodel,CBZmaterialfromsupplierIIIhadthe highestCBZDHcontent,whilethematerialfromsupplierIhadthe lowestCBZDHcontentamongstthethreecommerciallysupplied Table1
DHcontentintestsamplesascalculatedbyIAmodel,RamanmodelandcombinedRaman-IAmodel.
Testsamples DH[%(w/w)] Particles[#] DH[%(n/n)] DH[%(n/n)] DH[%(n/n)]
IAmodel Ramanmodel Combinedmodel
1 20w% 2250 17.1n% 19.3n% 18.7n%
2 50w% 3508 30.2n% 48.3n% 46.9n%
Fig.5.(I)CBZAH(red)andCBZDH(blue)classificationachievedbythecombinedRaman-IApartialleastsquarediscriminantanalysis(PLS-DA)model:Scorevaluesforthe first,secondandthirdlatentvariable(LV)oftheauto-scaledandmeancenteredcommerciallysuppliedCBZsamples.CBZSupplierI,n=3218(a),CBZSupplierII,n=3428(b) andCBZSupplierIII,n=3712(c).Theellipsoidsindicatethe95%confidenceintervals.(II)Visualobservationoftheparticleshapeandhomogeneityofcommerciallysupplied CBZmaterials(d,e,f)incomparisonwithpurerecrystallizedCBZsamples(g,h)wasperformedbyusingscanningelectronmicroscopy.
Table2
DHcontentincommercialsamplesasassessedbytheIA,Ramanandcombinedmodel.
Commercialsamples Particles[#] DH[%(n/n)]
IAmodel DH[%(n/n)] Ramanmodel DH[%(n/n)] Combinedmodel CBZSupplierI 3218 6.8% 6.1% 5.1% CBZSupplierII 3428 44.2% 64.9% 61.8% CBZSupplierIII 3712 23.0% 15.2% 14.6%
materials(Table2).Aspreviouslyconcludedbasedonaccuracy, theCBZDHcontentdeterminedbytheRamanandthecombined modelweresimilartoeachother(FigureS13,Supporting Infor-mation),whereastheCBZDHcontentdeterminedbytheIAmodel waslowerforthecommercialsampleIandhigherforcommercial sampleIIandIIIcomparedtotheRamanandthecombinedmodel results.Itshouldbenotedthattheappliedcutoffinparticlesize affectedtheabsoluteresults.
3.4. Criticalaspectsofmodeldevelopment
The accuracy of prediction of the resulting models with an increasingnumberofparticlesispresentedinFig.6.Thisaspectisan importantpartofmodeldevelopment,becauseitwilldirectlyaffect thetimeneededfortheanalyticalwork,aswellastherequired computingpower.Thefinalnumberofparticlesneededfora clas-sificationmodelwasestimatedtobereached,whenthepredicted amountofhydratewasstartingtovisuallylevelofftoaconstant value.Itwasobservedthatthenumberofparticlesleadingto sta-bleresultsisapproximately200particleswhenonlyIAorRaman modelswereused,whileforthecombinedmodeltheresults lev-eledoffalreadyataround100particles.Lessparticleswereneeded forastableresultwiththecombinedmodel,becausemore comple-mentaryinformationiscollectedfromeachparticle.Thisindicates thatcombiningtheinformationgainedbyvariouscharacterization techniquesisvaluablewithrespecttodetectionandquantification ofsolidformcomposition.
Particlesizeandshapeanalysisisanimportantanalyticaltask relatedtomanypharmaceuticalchallenges,however,this analy-sisisoftenperformedmanually,requiringalotoftimeandeffort [28].Today,however,thesemethodshaveadvancedtoalevelof automaticsampledispersionandparticletrackingsuitablefor auto-matedIAofhundredsofthousandsofparticleswithinalimited
timeframe.Imagingtechniquessuchasflowimagingmicroscopy havebeenappliedinanalyzingthepresenceofsmallparticlesin parenterallyadministeredpharmaceuticals[29].Anunsupervised model hasbeenappliedearlier for classification ofparticulates basedontheirmorphologybyusingneuralnetworkanalysis(ANN) [30].Thesemethodsareoftenreferredtoasa‘blackbox’approach, becausetheydonotallowmeansforcontrollingordeeper under-standingoftherootcauseresultinginthisclassification.Taking thisintotheaccount,whenanalyzingunknownpharmaceutical materials,thefollowing strategyis proposed.First, an unsuper-visedclassificationoftheparticulatemattercouldbeperformed basedonIAasthefastest andthemostaccessiblemethodfora largenumberofparticles.Theclassesobtainedbyunsupervised learningcanhelpintheprocesstoidentifyimpuritiesor inhomo-geneitywithinthematerials,i.e.thediscriminative particlesize andshapeparameterscanbeextractedrelatedtothe inhomogene-ity.Subsequently,morespecificanalyticalchemistrymethod,such asRamanspectroscopycanbeusedforchemicaland/orphysical identificationoftherootcauseforthediversityinthesample.If boththeselectedanalyticalchemistryapproachandtheIAofthe unknownspecimenareavailable,acombinedmodelcanbecreated forafastscreeningofthesematerialswithhigheraccuracyand pre-cision.Onecanusethisstrategytoscreenforinhomogeneitywithin pharmaceuticalmaterialsaswellasotherfinechemicals,to iden-tifytheiroriginbasedonidentificationofuniqueimpurityprofiles andfinally,tocreatefastmethodsforqualitycontrol.Byapplying individualparticleanalysisapproaches, itis theoretically possi-bletodetectsolidformimpuritiesatverylowlevels,ultimately atasingleparticlelevel.IAallowsforananalysisofahigh num-berofindividualparticlesandithasalowertheoreticalsensitivity comparedtoRamanspectroscopyatthecostofapproximately5 %uncertainty,i.e.5timeshighercomparedtothespectroscopic model.
Fig.6.CBZDHcontentoftestsamples1(dottedline),2(brokenline)and3(solidline),asaresultofusingtheRaman(a),IA(b)andthecombinedRaman-IAPLS-DAmodels (c).Modelsarecreatedbyusingdataofanincreasingnumberofrandomlysampledparticles.
4. Conclusion
Inthisstudythreedifferentclassificationmodelswere devel-opedforsolidformcharacterizationatasingleparticlelevel.The IAmodelwasthebestforfastscreeningandcomparisonofdifferent samplesaccordingtotheirhydratecontent.TheIAofatleast200 particleswassufficientforaclassificationerroroflessthan5%.This canberelevantforanumberofapplicationswheretimeandcost ofanalysisarethehighestpriority.TheRamanmodelcouldassess solid form compositionwith2 %classificationerror andwould thereforebetheidealchoicewhereaccuracyandprecisionabove98 %areneeded.WithacombinedmodelbasedonIAandRaman spec-troscopyonlyapproximately100particleswereneededtoassess thesolidformwith1%classificationerror.Thecombinationmodel hadthehighestprecision,sensitivityandspecificitycomparedto theIAmodelortheRamanmodelalone.IAandRamanspectroscopy arecomplementaryinassessingthesolidformdiversityof materi-alsandtheircombinationresultedinanaccuracyandprecisionof above99%.Thereportedapproachisdemonstratingthepotentialof integratedcomplementarymeasurementtechniquesinmeasuring verylowamountsofsolidformimpurities.
CRediTauthorshipcontributionstatement
AndreaSekulovic:Conceptualization,Methodology,Software, Validation,Formalanalysis,Investigation,Resources,Datacuration, Writing-originaldraft,Writing-review&editing,Visualization, Supervision, Project administration, Funding acquisition. Ruud Verrijk: Conceptualization,Resources, Writing - review & edit-ing,Projectadministration,Fundingacquisition.Thomas Rades: Conceptualization,Writing -review&editing.Adam Grabarek: Investigation,Writing-review&editing.WimJiskoot:Resources, Writing -review &editing. AndreaHawe:Resources,Writing -review&editing.JukkaRantanen:Conceptualization, Methodol-ogy,Resources,Datacuration,Writing -originaldraft,Writing -review &editing,Visualization, Supervision,Project administra-tion,Fundingacquisition.
AppendixA. Supplementarydata
Supplementary materialrelated tothis article canbefound, in the online version, atdoi:https://doi.org/10.1016/j.jpba.2020. 113744.
DeclarationofCompetingInterest
DrReddy’sIPDOLeiden(Leiden,Netherlands)hasfinancedthe PhDprojectofAndreaSekulovic.AndreaSekulovicandRuudVerrijk areemployedatDrReddy’s.JukkaRantanenandThomasRades havenotreceivedanyconsultingfeesforthiswork.
References
[1]D.J.W.Grant,Theoryandoriginofpolymorphism,in:Polymorphismin PharmaceuticalSolids,1999,pp.1–34.
[2]Q.Du,X.Xiong,Z.Suo,P.Tang,J.He,X.Zeng,Q.Hou,H.Li,Investigationofthe solidformsofdeferasirox:solvate,co-crystal,andamorphousform,RSCAdv. 7(68)(2017)43151–43160.
[3]F.Flicker,V.A.Eberle,G.Betz,Variabilityincommercialcarbamazepine samples-impactondrugrelease,Int.J.Pharm.410(1-2)(2011)99–106.
[4]R.Censi,P.DiMartino,Polymorphimpactonthebioavailabilityandstability ofpoorlysolubledrugs,Molecules20(10)(2015)18759–18776.
[5]A.Banerjee,J.Qi,R.Gogoi,J.Wong,S.Mitragotri,Roleofnanoparticlesize, shapeandsurfacechemistryinoraldrugdelivery,J.Control.Release238 (2016)176–185.
[6]J.D.Dunitz,J.Bernstein,Disappearingpolymorphs,Acc.Chem.Res.28(4) (1995)193–200.
[7]D.K.Bucar,G.M.Day,I.Halasz,G.G.Z.Zhang,J.R.G.Sander,D.G.Reid,L.R. Macgillivray,M.J.Duer,W.Jones,Thecuriouscaseof(caffeine)(benzoicacid): howheteronuclearseedingallowedtheformationofanelusivecocrystal, Chem.Sci.4(12)(2013)4417–4425.
[8]D.K.Bucar,R.W.Lancaster,J.Bernstein,Disappearingpolymorphsrevisited, Angew.Chem.Int.Ed.Engl.54(24)(2015)6972–6993.
[9]J.J.Maresca,Draftguidanceforindustry.ANDAs:pharmaceuticalsolid polymorphism.chemistry,manufacturingandcontrolsinformation,Journal ofGenericMedicines:TheBusinessJournalfortheGenericMedicinesSector2 (3)(2005)264–269.
[10]L.Y.Pfund,A.J.Matzger,Towardsexhaustiveandautomatedhigh-throughput screeningforcrystallinepolymorphs,ACSComb.Sci.16(7)(2014)309.
[11]H.M.A.DeArment,Celgenevs.Dr.Reddy’sRevlimidPatentInfringementCase DependsonPolymorphStability;LackThereofMayFavorCelgenePosition– Experts(accessed19August2019)https://www.drugpatentwatch.com/blog/ celgene-vs-dr-reddys-revlimid-patent-infringement-case-depends-polymorph-stability-lack-thereof-may-favor-celgene-position-experts/. [12]E.M.Paiva,V.H.daSilva,R.J.Poppi,C.F.Pereira,J.J.R.Rohwedder,Comparison
ofmacroandmicroRamanmeasurementforreliablequantitativeanalysisof pharmaceuticalpolymorphs,J.Pharm.Biomed.Anal.157(2018)107–115.
[13]A.Tulcidas,N.M.T.Lourenc¸o,R.Antunes,B.Santos,S.Pawlowski,F.Rocha, Crystalhabitmodificationandpolymorphicstabilityassessmentofa long-acting2-adrenergicagonist,CrystEngComm21(22)(2019)3460–3470.
[14]M.Maghsoodi,Roleofsolventsinimprovementofdissolutionrateofdrugs: crystalhabitandcrystalagglomeration,Adv.Pharm.Bull.5(1)(2015)13–18.
[15]B.W.Kammrath,A.Koutrakos,J.Castillo,C.Langley,D.Huck-Jones, Morphologically-directedRamanspectroscopyforforensicsoilanalysis, ForensicSci.Int.285(2018)25–33.
[16]N.Chieng,T.Rades,J.Aaltonen,Anoverviewofrecentstudiesontheanalysis ofpharmaceuticalpolymorphs,J.Pharm.Biomed.Anal.55(4)(2011) 618–644.
[17]E.A.deMoura,M.V.C.Terto,E.A.deMouraMendonca,J.V.Procopio,etal., Solid-stateformcharacterizationofriparinI,Molecules22(10)(2017).
[18]L.C.Lee,C.Y.Liong,A.A.Jemain,Partialleastsquares-discriminantanalysis (PLS-DA)forclassificationofhigh-dimensional(HD)data:areviewof contemporarypracticestrategiesandknowledgegaps,Analyst(Cambridge,U. K.)143(15)(2018)3526–3539.
[19]S.Favilla,C.Durante,M.L.Vigni,M.Cocchi,Assessingfeaturerelevancein NPLSmodelsbyVIP,Chemometr.Intell.Lab.Syst.129(2013)76–86.
[20]K.Kogermann,J.Aaltonen,C.J.Strachan,K.Pollanen,P.Veski,J.Heinamaki,J. Yliruusi,J.Rantanen,Qualitativeinsituanalysisofmultiplesolid-stateforms usingspectroscopyandpartialleastsquaresdiscriminantmodeling,J.Pharm. Sci.96(7)(2007)1802–1820.
[21]C.J.Strachan,P.F.Taday,D.A.Newnham,K.C.Gordon,J.A.Zeitler,M.Pepper,T. Rades,Usingterahertzpulsedspectroscopytoquantifypharmaceutical polymorphismandcrystallinity,J.Pharm.Sci.94(4)(2005)837–846.
[22]K.Maruyoshi,D.Iuga,A.E.Watts,C.E.Hughes,K.D.M.Harris,S.P.Brown, Assessingthedetectionlimitofaminoritysolid-stateformofa pharmaceuticalbyHdouble-quantummagic-anglespinningnuclear magneticresonancespectroscopy,J.Pharm.Sci.106(11)(2017)3372–3377.
