This article was downloaded by: [Xiaonan Lu] On: 10 July 2012, At: 18:08 Publisher: Taylor & Francis Informa Ltd Registered in England and Wales Registered Number: 1072954 Registered office: Mortimer House, 37-41 Mortimer Street, London W1T 3JH, UK Critical Reviews in Food Science and Nutrition Publication details, including instructions for authors and subscription information: http://www.tandfonline.com/loi/bfsn20 Determination of Antioxidant Content and Antioxidant Activity in Foods using Infrared Spectroscopy and Chemometrics: A Review Xiaonan Lu a & Barbara A. Rasco a a School of Food Science, Washington State University, Pullman, WA, USA Accepted author version posted online: 27 Jul 2011. Version of record first published: 02 Jul 2012 To cite this article: Xiaonan Lu & Barbara A. Rasco (2012): Determination of Antioxidant Content and Antioxidant Activity in Foods using Infrared Spectroscopy and Chemometrics: A Review, Critical Reviews in Food Science and Nutrition, 52:10, 853-875 To link to this article: http://dx.doi.org/10.1080/10408398.2010.511322 PLEASE SCROLL DOWN FOR ARTICLE Full terms and conditions of use: http://www.tandfonline.com/page/terms-and-conditions This article may be used for research, teaching, and private study purposes. Any substantial or systematic reproduction, redistribution, reselling, loan, sub-licensing, systematic supply, or distribution in any form to anyone is expressly forbidden. The publisher does not give any warranty express or implied or make any representation that the contents will be complete or accurate or up to date. The accuracy of any instructions, formulae, and drug doses should be independently verified with primary sources. The publisher shall not be liable for any loss, actions, claims, proceedings, demand, or costs or damages whatsoever or howsoever caused arising directly or indirectly in connection with or arising out of the use of this material. CriticalReviewsinFoodScienceandNutrition,52:853–875(2012) Copyright(cid:1)(cid:1)C TaylorandFrancisGroup,LLC ISSN:1040-8398/1549-7852online DOI:10.1080/10408398.2010.511322 Determination of Antioxidant Content and Antioxidant Activity in Foods using Infrared Spectroscopy and Chemometrics: A Review XIAONANLUandBARBARAA.RASCO SchoolofFoodScience,WashingtonStateUniversity,Pullman,WA,USA 2 1 Developingrapidanalyticalmethodsforbioactivecomponentsandpredictingboththeconcentrationandbiologicalavail- 0 2 abilityofnutraceuticalcomponentsinfoodsisatopicofgrowinginterest.Here,analysisofbioactivecomponentsandtotal y ul antioxidantactivityinfoodmatricesusinginfraredspectroscopycoupledwithchemometricpredictivemodelsisdescribed. J 0 Infraredspectroscopyoffersanalternativetowetchemistry,chromatographicdeterminationofantioxidants,andinvitro 8 1 biochemicalassaysforassessmentofantioxidantactivity.Spectroscopicmethodsprovideatechniquethatcanbeusedwith 0 biologicaltissueswithoutextraction,whichcanoftenleadtodegradationoftheantioxidantcomponents.Samplepreparation 8: 1 timegreatlydecreasesandanalysistimeisveryshortonceapredictivemodelhasbeendeveloped.Spectroscopicmethods at can have a high degree of precision when applied to analysis of nutraceutical compound concentration and antioxidant ] u activityinfoods.Thisarticlesummarizesrecentadvancesinvibrationalspectroscopyandchemometricsandapplications L n ofthesemethodsforantioxidantdetectioninfoods. a n o Xia Keywords Antioxidants,IRspectroscopy,chemometricmodels,food,plant,bioactive [ y b d e d oa INTRODUCTION ingcatalase,glutathioneperoxidase,andsuperoxidedismutase wnl (MinandBoff,2002).Recentliteraturesupportsthenotionthat Do The generation of reactive oxygen species is unavoidable antioxidantcomponentsinfoods,suppliedfromthedailydiet, during aerobic metabolic processes in human body cells and canquenchreactiveoxygenspecies,aidinthefunctionalperfor- the oxidation products that are generated through these reac- manceofenzymesystemsforself-defensemechanismswithin tions can initiate cell damage which is associated with aging cellsandtherebyreducetheriskofmanyhumandiseases(Choe and cancer. Reactive oxygen species include singlet oxygen, andMin,2009).Epidemiologicalstudieshavealsoreportedthe hydroxyl radicals, hydrogen peroxide and hydrochlorous acid, relevanceofantioxidativefunctionalfoods,andnutritionalsup- and peroxynitrite (Fang et al., 2002). Reactive oxygen species plements or nutraceutical compounds to maintain or improve cancauseDNAstrandbreakage,inducemodulationofgeneex- healthandreducetheriskorimpactofchronicdiseases(Kaur pression,andoxidationofproteinsandlipid(Leeetal.,2004). andKapooretal.,2001;Stanneretal.,2004). In the last decade, researchers have made substantial progress Numerousworksoverthepast20yearshavestudiedthean- investigating the role of reactive oxygen species and their as- tioxidantcapacityofvariousfoodcomponents.Manyanalytical sociation within many types of age-related diseases, including methods have been developed to determine concentrations of stroke,cardiovasculardisease,asthma,arthritis,retinaldamage, nutraceuticalcomponentsaswellastheantioxidantactivityfor chronic obstructive pulmonary disease, neurodegeneration, di- these components within food matrices (Dillard and German, abetes, and dermatitis (La Vecchia et al., 1998). The first line 2000). For nutraceutical compounds, high performance liquid of defense in the body is a series of enzyme pathways involv- chromatography(HPLC)istheprimarymethodforseparation, identification, and quantification (Hurst, 2002). Most recently, massspectrometryhasbeenincorporatedasadetectionmethod AddresscorrespondencetoBarbaraA.Rasco,P.O.Box646376,Schoolof coupled with liquid chromatography to satisfy the require- FoodScience,WashingtonStateUniversity,Pullman,WA99164-6376,USA. E-mail:[email protected] ment of many investigators for analytical methods with lower 853 854 X.LUANDB.A.RASCO detection limits (∼parts per billion levels). For determination regionandcanbedividedintothreesub-regionsbasedondiffer- of antioxidant activity, in vitro chemical or cell culture assays entwavenumberareas,namelythefar-,mid-,andnear-infrared arecommonlyutilized(Huangetal.,2005;Wolfeetal.,2008); region(Workman,1999).Themid-infraredregion(from4000to animalbioassaysarealsobecomingmorecommonasameans 400cm−1)providesmoreusefulvibrationalinformationabout to more accurately determining the effect of these compounds thefunctionalgroupsofmoleculesandthusitisusedfrequently on organ function and growth (Sivalokanathan et al., 2006) as for research in biomedical science (Jackson et al., 1997; Nau- well as clinical trials. However, scientists are striving to find mann, 2001), measurement of components in biological tis- alternative simpler and potentially more rapid techniques that sues (Movasaghi et al., 2008), detection of atmosphere pollu- couldcomplementorreplacetraditionalassaysparticularlyfor tants (Stoch et al., 2001), elucidation of protein and enzyme screeningfooditemsforqualityassuranceapplications.Infrared secondarystructures(Raussensetal.,1997),medicaldiagnos- spectrometryisonetechniquethathasgarneredinterestandhas tics (Petrich, 2001), and analysis of food components (Grif- advantageous features of a rapid analysis time, high precision fiths and Pariente (1988); Proctor et al., 2004; Hashimoto and and simple sample preparation methods, involving use of few Kameoka, 2008; Lucas et al., 2008; Lin et al., 2009; Subra- reagentsorsteps(Linetal.,2004).Thesefeaturesmakeinfrared manianetal.,2009;Kocaetal.,2010).Inthisspectralregion, spectrometryanexcellenttoolforscreeningfoodsforantioxi- organicmoleculesvibrateorrotateatspecificfrequenciescor- dantactivityandpotentiallyforquantifyingspecificcompounds respondingtodiscreteenergylevels.Tobe“infraredactive,”the within a complex matrix. Infrared spectroscopy is commonly moleculesmusthavechangesinthepermanentdipole(Nicolaou usedforroutineanalysisinfoodindustrybecauseoftheability andGoodacre, 2008).Forexample, thehomonuclear diatomic 2 toprocesslargesamplevolumesinashorttimeperiod(Rubio- molecules, such as hydrogen, nitrogen, and oxygen have zero 1 0 2 Diazetal.,2010). dipole moment; thus, no IR absorption is observed. However, uly Thisreviewprovidesanoverviewofthebasicknowledgeand heteronuclear diatomic molecules, such as hydrogen chloride 0 J principlesforbothmid-infraredandnear-infraredspectroscopy. and carbon monoxide possess a permanent dipole moment, so 8 1 Important factors for current analytical protocols of infrared they can absorb IR energy. However, the carbon dioxide is in- 8:0 spectroscopy such as the use of an attenuated total reflectance active in IR because the vibration produces no change in the 1 (ATR)cellfortestingofbiologicalmaterialssuchasfoodand dipole moment. Figure 1 illustrates how infrared active func- ] at foodingredientsandforthedataanalysiscurrentlyutilizedfor tional groups present in IR spectra can provide a great deal of u L Fouriertransformationanddevelopmentofchemometricbased analyticalinformationaboutafoodcomponent;informationthat an prediction models. Important factors associated with sample wouldotherwiseinvolvecomplicatedchromatographicanalysis n o preparation and data collection will be introduced including andcouldbedifficulttoobtain.HereanIRspectrumofacontrol a Xi handling biological samples and data preprocessing (i.e., bin- soybeanoiliscomparedtothatofaphoto-isomerizedoilcon- [ y ningandsmoothing)asappliedtorawspectrawillbedescribed. taining10%totalconjugatedlinoleicacidindicatingchangesin b d Apresentationofimportantchemometricmodelssuchaspartial CHstretch,carboxylatemoietiesandrelativeamountsoftrans- e d leastsquares(PLS)regressionmodelswillbediscussedinrele- transandtrans-cisisomersthatcanoccurduringisomerization a o nl vantdetailincludingthetheory,stepstoestablishthemodel,the (Fig. 1). Experiments could be conducted to monitor both the w factors affected model rigor, key parameters for model estab- extent and rate of oxidation, and oxidation products could be o D lishment(i.e.,determinationofthenumberoflatentvariablesin quantified, presuming analytical standards are available. Most apredictionmodel),andidentificationofanalyteswillbepre- importantly,thisfigureillustrateshowIRcanbeusedtomon- sented. Furthermore, what chemometric methods are and how itorlossofantioxidantsinafoodsystem,evenwhentheexact these can be developed and validated to estimate the concen- components responsible for antioxidant activity have not been trationsofnutraceuticalsandantioxidantactivityaccuratelyare fullyascertained. alsointroduced. Theapplicationofinfraredspectrometryinthispaperisdi- videdintotwopartsforidentificationandpredictionofthecon- PrinciplesofFourierTransformInfraredSpectroscopy centrationofbioactivecomponentsinfoodandnutraceuticals, antioxidants,andphytochemicalsinfoodcomponentscommon The development of applications of infrared spectroscopy inthedailydietwillbepresentedwiththecurrentliteratureson tofoodchemistryandappropriatedataanalysisrequireanun- analysisbyIRsummarized.Forantioxidantactivitydetermina- derstanding of the theory and principles involved with spec- tion, several selective food matrices will be chosen including troscopy. In the infrared region, the IR spectrum is composed berries, tea, and red wine and provided as a basis for discus- of innumerable infinite narrow bands of monochromatic light. sion of how IR methods can be used for other foods. Finally, In common instrument configurations beam splitters are used. suggestionsforfutureresearchwillbepresented. When monochromatic light arrives at beam splitter (no light absorption), 50% light is reflected to a fixed mirror back to GENERALPRINCIPLESOFINFRA-RED thebeamsplitter;theotherhalfpassesthroughabeamsplitter, SPECTROSCOPY arrives at a moving mirror, then reflects back and recombines with the former beam (Fig. 2) (Chalmers and Griffiths, 2002). Theinfraredregionisdefinedastheregionoftheelectromag- The different paths of these two beams are referred to as the neticspectrumbetweenthevisiblelightregionandmicrowave opticalpathdifference(OPD).Adetectorisusedtorecordthe INFRAREDSPECTROSOCOPYANDCHEMOMETRICS 855 2 1 0 2 y ul J 0 1 8 0 8: 1 at u] Figure1 ATR-FTIRspectraofthe4000–650cm−1regionof(A)controlsoybeanoiland(B)photo-isomerizedoilcontaining10%totalconjugatedlinoleicacid L n (CitedfromKadamneetal.,2009). a n o Xia interferenceorsuperpositionofthesetwobeams,withtheresult- [ ingspectrumreferredtoaninterferogram.Theinterferogram(I) y b isafunctionofOPD(Smith,1996): d e d a o I(δ)=B(ν)cos(2πνδ) nl w o D Theinterferogramequationpresentedhereisformonochro- matic light with the wave number of ν and provides the basis forFouriertransformspectroscopy.B(ν)standsforintensityof monochromaticlightofthewavenumberofν.I(δ)istheinter- ferogram(cosineFouriertransformationofB(ν)).Whenalight source is continuous, or composed of numerous frequencies, theinterferogrambecomesmorecomplicated.Foracontinuous lightsource,thesignalintensitycanbeexpressedas: (cid:1) +∞ I(δ)= B(ν)cos(2πνδ)dν −∞ whereI(δ)expressesthesignalintensityatthepointthatOPD isδ.Thissignal(I(δ))isthesummationofallwavenumbers(δ) intherealmof-∞to+∞.Becausetheδ ischangingcontinu- ously,acompleteinterferogramcanbereceived.However,the equation above provides only the interferogram, and the data obtained in the interferogram cannot be interpreted directly. Figure2 Asimplespectrometerlayout(CitedfromThermoNicoletCooper- One more step is necessary, and a Fourier inverse transforma- ation,2001). 856 X.LUANDB.A.RASCO 2 1 0 2 y ul J 0 1 08 Figure3 FT-IRinstrumentaloperation(Fouriertransformation). 8: 1 ] at tionisused to“decode” the signalintensityateach individual troscopy,andthesewerecommonlyforapplicationsinthenear u L wavenumber so as to make identification of individual spec- infra-redwithdrysamples.Aclassicalsamplepreparationtech- n tral features possible (Griffiths, 1992). This transformation is niqueinthemid-IRrequiredgrindingamaterialtoafinepow- a n o presentedhere: der and dispersing it into a potassium bromide (KBr) matrix a Xi andthiswasnotusuallyappropriateforfoodmaterialswiththe (cid:1) y [ +∞ exceptionofrelativelypure,dryingredients,andinevitablyre- b B(ν)= I(δ)cos(2πνδ)dδ d sultedinproblemswithreproducibility(Griffithsetal.,1986). e −∞ d TheintroductionoftheAttenuatedTotalReflectance(ATR)cell a nlo Figure 3 illustrates the detailed process of Fourier transfor- overcamemanyoftheobstaclesofanalyzingfoodmaterialsby w infraredspectroscopy(WinderandGoodacre,2004).ATRisar- o mation. There are many advantages of Fourier transformation D guablythemostcommondetectionsystemforwidespreaduse including:theFellgettadvantage(ortheabilitytodetectmultiple forFT-IR(Milosevic,2004).Useofthistechniquereducesthe wavenumbersimultaneously),theJacquinotadvantage(higher complexity of sample preparation and measurement of intact optical throughput compared to other types of spectroscopy), biological samples, minimizing tissue damage to the greatest andtheConnesadvantage(theabilitytouseaninternalwave- extent possible.Sample extraction isnotusuallyrequired,and lengthcalibration).Burgulaetal.(2007)andSubramanianand forthemostpart,anyfoodcanbeanalyzedaslongasenough Rodriguez-Saona(2009)providedadetailedsummaryofthese water isremoved toreduce interference ofwater withspectral advantages applied in food science research. In general, the features of lipid, protein, and carbohydrate. The ATR acces- overallobjectiveofFT-IRistoimprovethesignal-to-noiseratio sory with many FT-IR instruments is commonly made of zinc (SNR)toincreasesensitivityandprovideaspectroscopictech- selenide crystal, but sometimes, germanium or silicon crystals niquethatisbothaccurateandreproducible(MarkandGriffiths, may be used (Milosevic and Berets, 2002). The principle un- 2002;Shaoetal.,2002).Coates(1999)reviewedthemathemat- derlying the operation of an ATR is as follows. When a beam icalmethodsforeliminationofbroadbackgroundinterferences oflightislaunchedwithanincidentangleofα fromtheinside frominfraredspectra. 1 to the surface of the crystal, the incident angle (α ) is smaller 1 than the refraction angle (γ ) because the refractive index of 1 AttenuatedTotalReflectance(ATR)anditsRoleinIR the crystal is larger than that of air (which is arbitrarily des- Analysis ignated as one). Along with the increase of α , γ will also 1 1 increase proportionately. When α increases to the critical an- 1 Before 1993, the only infrared techniques appropriate for gle, α2, the refraction angle is equal to 90 degrees, and at this foodscienceresearchweretransmittanceandreflectancespec- pointtherefractedlightwillbetransmittedthroughthesurface INFRAREDSPECTROSOCOPYANDCHEMOMETRICS 857 be collected by an attenuation of the energy of the evanescent γ wave.Inregionsoftheinfraredspectruminwhichthesample 1 absorbs energy, the evanescent wave will be attenuated. The vibration amplitude (intensity of evanescent wave) will decay α 1 exponentiallyfromthesurfaceofthecrystalandfinallydisap- pear(ObergandFink,1998).Whenthevibrationalamplitudeof theevanescentwavedecaysto1/eoftheoriginalincidentone, (1) Reflectance & Refraction (a1) thedistanceisreferredtoasthepenetrationdepth. The penetration depth depends upon the wavelength of the incidentlight,therefractionpropertiesofthecrystalandofthe γ 2 sample,andtheincidentanglebetweentheimpinginglightand thesurfaceofthecrystal.OneoftheproblemswefaceusingATR α 2 inthemid-IRregionforfoodsanalysisisthefactthatthereisone order of magnitude difference between the penetration depth into the sample of the radiation at higher frequencies (lower wavenumbers)thanatlowerfrequencies(higherwavenumbers) (2) Critical angle (a2) overthisspectralrange.Thisresultsintheintensityoftheab- sorbancepeakatlowfrequenciesbeingmuchhigherthanatthe 2 higherfrequencies(Hebertetal.,2004).Therefore,acorrection 1 0 2 in ATR spectra is needed and can be conducted either before uly α3 orafterspectracollection(Coates,1998)tocompensateforthis 0 J effect. 8 1 In some cases, a liquid sample (∼20 µL) can be applied 8:0 (3) Total reflectance (a3) directly onto the surface of an ATR cell (Patz et al., 2004; 1 Versarietal.,2010).However,extensivedataprocessingwork ] at Figure4 Graphicrepresentationoftotalreflectance. needs to be done comparing spectra from liquid samples with u L thoseofsolidsamplestoensurethatmeasurementartifactsare an ofthecrystal.Whentheincidentangleα isgreaterthanα ,the notbeingincorporatedintochemometricmodels.Insuchcases, n 3 2 ao radiationwillbetotallyreflected(Fig.4). an experimental determination needs to be made whether it is Xi Figure 5 shows light paths in an ATR crystal. The material moreefficienttoincreasesamplepreparationtimetodehydrate [ y to be measured is placed on top of the crystal cell. When the samplesortodevelopamoreinvolvedcomputationalmodelto b d incidentangleisgreaterthanthecriticalangle,totalreflectance compensateforwaterabsorbance. e ad willoccurwithnoinfraredlightpenetratingthroughthecrystal o nl surface and entering the test sample. However, if there is no w o infrared light passing through the surface of the crystal, how TheoryofNearInfrared(NIR)Spectrometry D doestheinfraredlightinteractwiththesample?Wheninfrared light is reflected, it produces standing waves near the surface NIR spectroscopy utilizes the spectral range from 780 to ofthecrystalcalledan“evanescentwave”(Herminghausetal., 2500nm(12,500–4,000cm−1)andprovidescomplexstructural 1994). When the sample touches with the outer surface of the information related to the vibration behavior of combinations crystal, the evanescent wave will protrude only a few microns of bonds (Polesello and Giangiacomo, 1983). The NIR region (0.5 µ - 5µ) beyond the crystal surface and penetrate into the of the electromagnetic spectrum involves the response of the sampleateveryreflectancepoint(Fig.5).Theabsorbancecan moleculebondsO-H,C-H,andN-H.Thesebondsaresubjectto vibrationalenergychangesatNIRfrequencies,andtwovibra- tion patterns exist for these bonds, a stretching and a bending evanescent wave vibration.Theenergyabsorptionofthesemolecularvibrations crystal cell sample isreflectedinanabsorptionspectrum(BouveresseandMassart, 1996).InNIRspectroscopy,reflectedortransmittedradiationis measured.Thespectralcharacteristicsofanyparticularmaterial is affected by wavelength dependent scattering and absorption processes and is dependent upon the chemical composition of theproduct,aswellasonitslightscatteringpropertiesrelated tothemicrostructure,particularlyforbiologicalmaterials(Lin etal.,2008). The NIR region is divided into short-wave NIR (SW-NIR) Figure 5 Horizontal ATR accessory with representation of an evanescent wave. andcommonNIRatthewavelengthof1300nm.TheSW-NIR 858 X.LUANDB.A.RASCO regionisdefinedastheabsorptionbandofhighovertones,while detectorillustratinghowtheconfigurationofthisspectrometer NIRinvolvesthefirstorsecondovertones(Huangetal.,2003). differsfromoneinthemid-IRrange. Theabsorptionintensitywilldecreaseathigherovertones;there- fore,SW-NIRtendstobelesssensitiveandiscommonlyusedin the transmittance mode, although reflectance is also common. SW-NIRhasalongerpathlengthmakingitausefulanalytical method for bulk properties of materials such as moisture, fat, SAMPLEPREPARATIONANDSPECTRACOLLECTION and protein in agricultural products (Lin et al., 2004) and for predictingtheconcentrationofmacronutrientcomponents.NIR For the FT-IR, sample preparation is relatively simple and islessusefulforanalysisofmicronutrientsorforanycomponent no derivatization is necessary. The preparation steps may in- presentatlevelsofapproximately0.5%orless.Thesensitivity cludeextractionofspecificbioactivecompoundfromamatrix of NIR analyses is approximately 0.1% (Cen and He, 2007). byfiltration,milddehydrationofsamples,andapplicationofthe However,thetechniquecanbeusefulforpredictingthestability analytetoafiberorfilter,suchasanaluminumoxidemembrane of micronutrient components, for example, susceptibility of a filter(Linetal.,2005).Thesesamplepreparationstepsimprove material to lipid oxidation, oxidative browning, or enzymatic theintensityofthesignal,improvesthesignal-to-noiseratio(De deterioration when these factors can be tied to concentrations Nardoetal.,2009),anddecreasestheincidenceofoverlapping of moisture or lipid. There are a number of recent publica- peaks at the same wavenumber regions from compounds with tions summarizing current NIR techniques and applications to chemicallysimilarstructures(i.e.,differentfattyacidsorsteroid 2 foodqualityandsafety(Scotter,1997;JhaandMatsuoka,2000; hormonescanhaveoverlappingspectralfeaturesataroundthe 1 20 Ellis and Goodacre, 2001; Cen and He, 2007; Nicolai et al., wavenumberof3000cm−1).Amilddehydrationstepdecreases uly 2007; Jimare Benito et al., 2008; Sun, 2009), and pharmaceu- theimpactofspectralfeaturesof“freewater”(Luetal.,2010) 0 J tical analyses and quality control (Aldrich and Smith, 1999; exhibiting a large peak around 3300 cm−1 (O-H stretching of 8 1 Reich,2005;Roggoetal.,2007;Gowenetal.,2008).Because water)andaround1700cm−1(O-Hanti-symmetricstretchingof 8:0 of the widespread use of NIR in industry, and with the analy- water).Thewaterspectracanmaskthemostimportantpeaksde- 1 sis of bulk commodities which can be important sources of or rivedfromotherbioactivecompoundsinfoodmatrices(Luand ] at carriers for nutraceutical components such as food oils, flour, Rasco,2010).Forthedehydratedsample,multiplespectraneed u L andsoymealorprotein,newapplicationswilllikelybedevel- to be obtained, and an average taken, with this value used for an opedforpredictionofbiologicallyactivecomponentsandtheir furtherPLSmodelestablishment(describedinalatersection). n o stability. Theselectionofthematrixformountingthedehydratedsample a Xi ConventionalNIRiscommonlyusedinthediffusereflection forIRdetectionisaconcern,withthemostcommonbeingalu- [ y mode.Hydrogenbondingfeaturescanbeobservedinthisregion minumoxidemembranefilter(Linetal.,2007)whichprovides b d ofthespectra.AschematicofanNIRspectrometerisshownin the least amount of interference, hydrophobic grid membrane e d Fig.6,includinglightsources,beamsplitter,system,reflector, filter(Mannigetal.,2008;Grassoetal.,2009),glassfiberfilter a o nl sample chamber, diffuse reflection detector, and transmission disks (De Nardo et al., 2009), and cellulose ester membrane w filter(Burgulaetal.,2006).Glassslidescanalsobeusedwhen o D there is no need to concentrate the analytes in the sample by filtrationpriortoanalysis(Luetal.,2010). Generally for NIR measurements, no sample preparation is neededexcepttoensurethattheparticlesizefordrysamplesis consistent from one sample to the next so that light scattering remains consistent (Cen and He, 2007). If samples need to be ground, care must be taken not to affect the moisture content duringsamplepreparationandstorageprocedures(Huangetal., 2003).BecauseofthesimplicityofsamplepreparationNIRis widelydeployedinfieldstudiesandindustryfacilitieswithout strict environmental control (Bouveresse and Massart, 1996). There are a number of rugged instruments available for field usewithdifferentdetectorconfigurationsandsamplechambers available for NIR spectrometers, for example, glass or quartz cuvettesofdifferentsizeforliquidsamplesmonitoredintrans- mittance,andmeasurementcellsforsolidsofdifferentsizesand configurations formeasurements indiffusereflectance. Useof Figure 6 Sketch of an NIR spectrometer. 1- light source, 2-beam splitter FT-IRinthefieldispossible,butremainsachallengesincemost system,3-reflector,4-samplechamber/detectorinletvalve,5-diffusereflection instruments are not designed to operate in environments with detector,6-transmissiondetector,7-controlanddataprocessinganalyzedsys- vibration,dust,orhighmoisturelevels. tem,8-printer(CitedfromCenandHe,2007). INFRAREDSPECTROSOCOPYANDCHEMOMETRICS 859 Table1 FrequencyofbandassignmentsforbiochemicalfeaturesFT-IR Theprocessesof“automaticbaselinecorrection”and“nor- spectra∗ malizing”therawspectramakecomparisonofspectrafeatures Frequency(cm−1) Assignment mucheasier.However,thesetechniquesmustbeusedproperly and conducting automated baseline correction routines can do ∼3600 O-Hstretchofhydroxylgroups more harm than good for quantitative analysis if there is no ∼3150 N-Hstretch ∼2960 C-Hasymmetricstretchof-CH3 understandingonthepartoftheanalystwhythesecorrections ∼2929 C-Hasymmetricstretchof>CH2 wouldbeappropriate(LuandRasco,2010). ∼2870 C-Hsymmetricstretchof-CH3 An “automatic baseline correction” is commonly first per- ∼2850 C-Hsymmetricstretchof>CH2 formed on the raw spectra to flatten baselines (Al-Qadiri et ∼1740 >C=Ostretchofesters al., 2008). Secondly, a “normalization” may be performed to ∼1650 AmideIofα-helix ∼1540 AmideIIofβ-sheet compensatefordifferencesinsamplethickness(Al-Holyetal., ∼1470 C-Hdeformationof>CH2 2006;Al-Qadirietal.,2006).Thesetwostepsforpreprocessing ∼1395 C=OsymmetricstretchofCOO− theFT-IRrawspectraareimportant,becauseduringtheacqui- ∼1305–1245 AmideIIIofproteins sition of ATR spectra, sample thickness can change to some ∼1240 P=Oasymmetricstretchof>PO2− extent (Al-Qadiri et al., 2006). For instance, a thicker mate- ∼1200–900 C-O,C-Cstretch,C-O-H,C-O-Cdeformation ∼1080 P=Osymmetricstretchof>PO2− rial can exhibit higher absorption than a thinner one, resulting in greater peak heights and peak areas, and sometimes an ac- ∗References: Naumann, 2001; Maquelin et al., 2002; Lu and Rasco, 2010; companying peak shift to a higher or lower wavenumber, for 2 Movasaghietal.,2008 example,becauseofthedifferentsaltconcentrationsinafood 1 0 2 matrix (Huang et al., 2001). Furthermore, measurement errors uly BANDASSIGNMENTS duringspectralcollectionmaycausethebaselinetotilt.With- 0 J out correcting the baseline, comparing spectra and conducting 8 1 Therearenumerouspublicationssummarizingbandassign- quantitativeanalysiswouldnotbepossible. 8:0 mentsinthemid-andnear-infraredregion.Mid-infraredband Otherdatapreprocessingmethods,suchasbinning,smooth- 1 assignmentismoreimportantduetoextensiveavailableinfor- ingfollowedbysecondderivativetransformation,magnifymi- ] at mationofmolecularstructure.Movasaghietal.(2008)reviewed nor differences among IR spectra (Goodacre, 2003). Binning u L numerous studies providing a database on the most important reducesthenumberofdatapointsinaspectrumbynpointsinto an mid-infraredcharacteristicpeakfrequenciesfornaturaltissues onepointandeliminatestheopticalimbalanceproblemassoci- n o analysisinthemid-IRregion.Otherrecentpublicationsprovide atedwithmanyarraybasedspectrophotometers(Al-Qadirietal., a Xi compilationsonimportantfrequenciesforfoodanalysis(Nau- 2008).Smoothingeliminateshighfrequencyinstrumentalnoise [ y mann,2001;Maquelinetal.,2002;Burgulaetal.,2007;Luand by averaging adjacent data points (Al-Holy et al., 2006). Sec- b d Rasco,2010).Table1providesashortlistofbandassignmentfor ondderivativetransformationseparatesoverlappingabsorption e d themostimportantIRabsorbancefeaturesforbiologicalsam- bands,eliminatesbaselineoffsets,increasestheapparentspec- a nlo plematrices.ItiscriticaltorememberthatIRabsorbancebands tralresolution,andprovidesanestimateofthenumberofover- w canonlyreflectinformationofmolecularfunctionalgroupsand lappingbandswithinaspectralregion(Al-Qadirietal.,2008). o D do not provide an identification of a specific chemical com- pound. Therefore, it is important to be knowledgeable of the contents of the infrared library and how this information can PARTIALLEASTSQUARES–MULTIPLELINEAR assistananalysttodeterminesamplecomposition.Thefirstpri- REGRESSIONMODELS ority is to determine and then justify band assignments in the targeted compounds, and secondly to understand how matrix Partial Least Squares (PLS) model is a multivariate regres- compositioncouldaffectthepositionofabsorbancebandsand sionmethodcommonlyusedtoestablisharelationshipbetween the subsequent interpretation of IR absorbance spectra so that referencevaluesforattributessuchascellnumbersorconcentra- spectraldataarecorrectlyinterpreted. tionofaparticularanalyte,andpredictedvaluesforthatattribute inatestsamplebaseduponitsinfraredspectralfeatures(Geladi andKowalski,1986;Al-Qadirietal.,2008).Itisanefficientsta- DATAPREPROCESSING tisticalpredictiontechnique,especiallysuitabletosmallsample data with many correlated variables (Alsberg et al., 1998). To Datapreprocessingalgorithmsareusefultoolstoenhancethe establish a PLS model, the first step is to choose the optimal spectraldifferencesbetweensamplesandtheuseofoneormore number oflatentvariables (Huang etal.,2001). Loading plots ofthesealgorithmsisoftenaprerequisiteformultivariatedata in the PLS model can be developed to justify a selection of a analysisroutines(JavisandGoodacre,2005).Itisimportantto smallnumberoforthogonalfactors(knownaslatentvariables) understandwhateachofthesepreprocessingtechniquesdo,and forconstructionofaPLSmode.Theseloadingscorrelatetothe nottooverusethemtocompensateforcollectionofpoorquality principalcomponentswithinadefinedwavenumberregionthat spectra. account for the greatest difference between samples in a data 860 X.LUANDB.A.RASCO set.Usingtoomanylatentvariablesdecreasestheprecisionof the model due to data over-fitting. Too low a number of latent variables will reduce the utility of the model since not all of therelevantdatahasbeenincorporatedintoforitsconstruction (Henningssonetal.,2001).Duringthereviewofthemanypub- licationsinthisarea,wefoundmajorproblemsinanumberof publishedworkswithoverfittingdata,specificallywithselection of a number of latent variables too great to justify chemically, casting doubt on the validity of the chemometric models pre- sented.Latentvariablescanbeassignedbaseduponweighting of spectral changes at specific wavelengths that contribute to thechemometricmodel,butwithoutacredibleassignmentofa chemicalfeaturetoalatentvariable,amodelisweak.Models with latent variable values higher than 10 are suspect unless a good understanding of the composition of the matrix and an- alytes within it exists and can credibly provide an explanation forthebasisforthemodel.Toomanylatentvariablesoverfitthe data,sonomatterhow“bad”thecollectedspectraare,increas- 2 ingthenumberoflatentvariablevaluetosomeextent(i.e.,latent 1 20 variable = 15) will always provide a “good” linear regression uly model.Specifically,ifthenumberoflatentvariablesarehigher 0 J than10forbiologicalexperiments(i.e.,biology,pharmacy,and 8 1 foodscience),thePLSmodelislikelynotreliable;areasonable 8:0 numberoflatentvariables(vectornumbers)isaroundto5to8 1 (BoulesteixandStrimmer,2006;YangandRen,2008). ] at The following factors are important if a robust calibration u L models is to be established: 1) having a sufficient number of an spectra,2)havingspectraforsamplesthathaveanalyteconcen- n o trationsevenlydistributedovertheentirerangeofinterest,and a Figure7 Cross-validated(leave-one-out)PLSplotsforlycopenecontentin Xi 3)aspectrallibraryforverifyingpeakassignmentsareimpor- tomatojuiceusingthedirectmethod(A)andthelipidextractionmethod(B) [ y tantfordevelopingachemometricmodelthatwillbeusefulfor (CitedfromDeNardoetal.,2009). b d predictingconcentrationofananalytefromspectralfeaturesof e d acomplexbiologicalmaterial,forexample,theantioxidantcon- a sociated with PLS models include a predicted residual sum of o nl tentortotalpathogencountsinafoodmatrix(StahleandWold, squares and a correlation coefficient (R) values. The objective w (1988). For a PLS correlation, reference data is needed from o istooptimizethemodelandimprovethecorrelation.Following D whichcorrelationsbetweenreferencevaluesandthepredicted cross-validation, the model is used to predict the concentra- valuescanbedeterminedbaseduponspectralfeatures(Fig.7). tionofananalyteinapreviouslyuntestedsample.Thesamples The standard error of prediction (SEP) is the most commonly selectedforthecalibrationandthevalidationsethavetobein- usedparametertocalculatethepredictiveperformanceofaPLS dependent;andpreferablytheyshouldconsistofsamplesfrom calibrationmodel. differentbatchesofmaterial,takenatdifferenttimes(Boulesteix andStrimmer,2006). TherearethreestepstoestablisharobustPLSmodel,namely PLSMODEL–STANDARDPROCEDURES calibration,cross-validation,andprediction.TheRvalueofthe cross-validationmodelisalwayslowerthanforthecalibration The optimal number of PLS latent variables to use in the modelandthestandarderrorofcross-validation(SECV)which PLS models is obtained by the cross-validation method with ishigherthanstandarderrorofcalibration(SEC).Forprediction, theobjectiveofselectinglatentvariablestoobtainthesmallest it is important to evaluate two parameters if the model is to rootmeansquareerrorofpredictionvalues(Zhangetal.,2004; provideagoodresult—standarddeviationandthecoefficientof Lin et al., 2008; Liu et al., 2010). Two cross-validation meth- variability(Martens,2001). odsarecommon:(1)leave-one-outand(2)atrainingmodelin ThereareadvantagesanddisadvantagesforPLSmodels.A which75%ofthetotalsamplesareselectedforcalibrationand majoradvantageis,afteraregressionmodelhasbeenestablished the remaining 25% for cross-validation. With a limited sam- and validated, analyte concentrations for a new sample can be plenumber,(1)isalmostalwaysthebestoption;otherwise(2) predictedquickly,usuallywithin5min,includingspectralcol- providesabettermethodforcrossvalidation,butismoretime lectionanddataconversion.Thisprovidesforaveryconvenient consumingandrequiresagreatersamplenumber.Statisticsas- andhighlyefficientanalyticalmethodinmanysituations,such INFRAREDSPECTROSOCOPYANDCHEMOMETRICS 861 astheonesfacedinqualitycontrolinthefoodindustry,where helptoimprovetheutilityofIRtechnology.Moredetailsabout numerous samples need to be processed quickly and there is transferofmultivariatemodelshasbeenrecentlyreviewed(Cen neitherthetimenorresourcestoruntheconventionalreference andHe,2007). method.However,therearedisadvantagestospectralanalyses aswell.First,infraredspectrometryisonlysensitivetoacertain componentsandoveralimitedrangeofconcentrations.Mid-IR APPLICATIONSOFIRMEASUREMENTSTO canquantifyanalytesinthelowpartperthousandrangeandNIR DETERMINECONCENTRATIONANDPREDICTTHE onlyto0.1%.Secondly,thePLSmodelmustbebuiltbasedupon BIOLOGICALACTIVITYOFFOODCOMPONENTS samples that contain a range of analyte concentrations. These samplesareoftendifficulttoobtain,andastheanalyteconcen- In general, these applications can be separated into the fol- trationchanges,spectralfeaturesofthematrixmayalsochange lowing divisions: (1) qualitative and quantitative analysis of a in a manner that is not necessarily easy to predict and which specific bioactive compound (2) rapid determination and pre- may be difficult to compensate for in the predictive model. A dictionofacategoryofbioactivecompounds(i.e.,polyphenols) simpleexampleismoistureinadriedmaterial.Asthemoisture and(3)rapiddeterminationandpredictionofantioxidantactiv- contentdecreases,alossinvolatileconstituentsmayoccur,lipid ity.NIRismainlyusedfor(2)and(3)andMIRisextensively oxidationandbrowningreactionproductsmayincreaseincon- employedfor(1)and(3).MIR,especiallyFT-IR,hasbeenused centration,allchangingthe“background”spectraforthematrix. toidentifyspecificcompoundsandforqualitativeandquantita- Althoughwehavemuchexperiencewithmoisturemodelsand tivedeterminationsofantioxidantsinfoods. 12 cancompensatefortheseeffectsfromdecadesofexperienceand However, no matter whether NIR or MIR are used for de- 0 2 millionsofspectraongrains,oilseed,andmeat,thesechanges termination of bioactive compounds or nutraceuticals, an ap- uly wouldbemuchmoredifficulttoaccountforinmodelsinvolving propriatereferencemethodneedstobeavailableandreference 0 J changesinthetotalcontentandrelativedistributionofantiox- assaysconductedatthesametimeasspectralmeasurementsif 8 1 idant components in plant tissue as a result of storage or pro- quantitationisanticipated.Infraredspectrometryisanindirect 8:0 cessing.Researcherswillbeabletoimprovethesophistication methodforquantificationandreferencemethodssuchasHPLC ] at 1 ofefatthuereirsmofodtheelsmasataricgerseacthearnugnededrusrtianngdipnrgocoefsshionwg itsheobsptaeicnterdal, abrieoaocfttievneccoomndpuocnteendtsto.Tphreonv,idaePqLuSanmtiotdateilvceoiunlfdorbmeaetsitoanblaisbhoeudt u L butthisisanimportantconsideration,andonethatresearchers andpredictionofconcentrationsofeitheraspecificcompound an shouldnotloosesiteof.Extrapolationbeyondtheconcentration ortotalantioxidantcapacitycouldbeachieveddependingupon n o rangeforwhichtherearereliablereferencevaluesisalsoarisk howthemodelisconstructed. a Xi andonethatwehaveobservedinourexaminationofpublished A number of the important bioactive nutraceutical com- [ y work.Third,thepredictivemodelneedstobecheckedroutinely pounds and antioxidant activity will be discussed in the fol- b d to account for optical shifts with spectral calibrations updated lowing sections, focusing on comparisons of parameters and e ad frequentlytokeepthemodelaccurate.Thisistimeconsuming techniques for analysis of each component (Table 2). It is im- o nl and laborious, and requires a continuous source of biological portanttoknowthatwhileallfoodsarefunctionalinthatthey w materialwithawiderangeofanalyteconcentration;however,it providenutrients,foodsthatcontain“nutraceuticals”allegedly o D iscriticalifchemometricmodelsaretoremainreliable.Finally containhealthpromotingcomponentsinadditiontonutrients.In andmostimportant,duringtheestablishmentofthePLSmodel, thecurrentreview,onlythe“smallmolecule”nutraceuticalswill operational conditions, all the parameter settings (i.e., prepro- be discussed. The “larger molecule” nutraceutical compounds cessingsteps),andthemeasurementtemperaturearerequiredto suchasdietaryfiberandbioactivepeptidesarenotcovered. bestandardizedandkeptconstantsincethesefactorswillaffect Before we go into details of each antioxidant compound thereliabilityofthespectraldataandtheresultingrigorofthe in various foods, we will use two examples to explain how analyticalmodel. infrared spectroscopy can be coupled with chemometrics in a Recently,somepublicationshavefocusedonacomplimen- stepbystepanalysis.Becausethegeneralstepsforanalysisof tary topic after the model established. It is called “transfer of nutraceuticalcomponentsinfoodsbythistechniquearesimilar, multivariate models” (Feudale et al., 2002). Multivariate cali- arepresentativeexamplewillhelpreadersunfamiliarwiththese bration models play a key role in the analytical measurement. methodstounderstandtheprincipleandthenbeabletousethis ThereliabilityofmeasurementfrombothFT-IRandNIRmeth- informationtodesignandperformspecificexperiments. odsdependsuponthecalibrationmodelusedandselectioncan varywithinstrumenttypeandinstrumentperformance.Theap- plication will be promising if the same spectral data set could Example1:UseofFT-IRwithPLStoDetermineandPredict beusedindifferentenvironmentsorinstruments(Shenketal., theContentofBioactiveConstituentsinFoods 1985).ThecalibrationmodelbuiltbyNIRScanbeusedtoan- alyze a large amount of samples on-line. Thus, sharing model Figure8showstheflowchartofstepsofthistechnique.For librariesandmakingitpossibletotransfermultivariatecalibra- eithercalibrationorcrossvalidationPLSmodel,thereference tion models and revalidate those on different instruments can value collection and the spectral feature collection need to be