One of the fields of research o f our laboratory is the use of Raman spectroscopy to study the chemical com- position of biological samples in view of its application as an in situ and in vivo detection method o f human diseases. Atherosclerosis is the main cause o f death in industrial countries. Therefore, any contribution to the understand- ing of the pathogenesis o f atherosclerosis is potentially o f major interest. Most o f the studies on atherosclerosis show the importance o f lipids in the disease pathogene- s i s . 1-6 However, current techniques o f investigation re- quire the r e m o v a l o f tissue samples and their destruction. In these cases, a nondestructive technique that could rap- idly provide biochemical information using very small amounts of tissue sample with no previous preparation would be of great relevance. R a m a n spectroscopy is an analytical technique based on the interaction of an incident m o n o c h r o m a t i c radiation with vibrational energy levels o f molecules. This tech- nique, applied in our laboratory for m a n y years in c o m - bination with optical fibers, allows in situ analyses giving both qualitative and quantitative information for various chemical mixtures and, especially, biological samples. 7-1~ In these cases, a near-infrared excitation source allowing fluorescence-free measurements should be used in order to obtain high-quality R a m a n spectra.
Quantitative Analysis of Cholesterol and Cholesterol Ester Mixtures Using Near-Infrared Fourier Transform Raman Spectroscopy P L e C A C H E U X , G MENARD, and N G U Y E N Q U Y D A O * H NGUYEN QUANG, P W E I N M A N N , M JOUAN, Laboratoire P.C.M., URA D1907 CNRS, Ecole Centrale Paris, 92295 Chatenay-Malabry Cedex, France Near-infrared Fourier transform Raman spectroscopy is a rapid and nondestructive technique that can provide reliable qualitative in situ information about the chemistry of biological samples W e have combined this technique with partial least-squares (PLS) regression to perform a quantitative determination of free and esterified cholesterol in two synthetic sample series In 66 ternary mixtures containing various proportions of cholesterol, cholesterol linoleate, and oleate, the standard errors of prediction were 1.27, 1.17, a n d % , respectively For the second series of experiments concerning the sensitive problem of quantitative analysis of cholesterol palmitate and stearate mixtures, the standard error of prediction for 49 samples was 3.02% It is also possible to extract quantitative information for a single component of the ternary mixtures independently These results are of great importance w h e n - - a s in the case of arterial s a m p l e s - - m a n y chemical species are present Comparison between Raman spectra of ternary mixtures and atherosclerotic rabbit aorta shows that many bands, assignable to free and esterified cholesterol, are easily observed in the aorta spectrum Index Headings: Lipids; Quantitative analysis; Raman spectroscopy; Partial least-squares regression; Cholesterol; Cholesterol esters INTRODUCTION One of the fields of research of our laboratory is the use of Raman spectroscopy to study the chemical composition of biological samples in view of its application as an in situ and in vivo detection method of human diseases Atherosclerosis is the main cause of death in industrial countries Therefore, any contribution to the understanding of the pathogenesis of atherosclerosis is potentially of major interest Most of the studies on atherosclerosis show the importance of lipids in the disease pathogenes i s 1-6 However, current techniques of investigation require the removal of tissue samples and their destruction In these cases, a nondestructive technique that could rapidly provide biochemical information using very small amounts of tissue sample with no previous preparation would be of great relevance R a m a n spectroscopy is an analytical technique based on the interaction of an incident monochromatic radiation with vibrational energy levels of molecules This technique, applied in our laboratory for many years in combination with optical fibers, allows in situ analyses giving both qualitative and quantitative information for various chemical mixtures and, especially, biological samples 7-1~ In these cases, a near-infrared excitation source allowing fluorescence-free measurements should be used in order to obtain high-quality R a m a n spectra Received 19 July 1995; accepted 22 M a y 1996 * Author to w h o m correspondence should be sent Volume 50, Number 10, 9 However, analyses of biological samples by R a m a n spectroscopy present us with two main problems On the one hand, the complexity of these samples leads to a great complexity in their Raman spectra Therefore, it is not easy to extract relevant information on specific components of interest On the other hand, the components to be studied can be of very similar structures Therefore, R a m a n spectra of the mixtures show overlapped bands (cases of long saturated hydrocarbon chains), making quantitative analysis very difficult In these cases, classical quantitative spectral analysis methods using peak intensity information are inaccurate However, it has been shown that partial least-squares (PLS) regression, which is a statistical multivariate method using factor analysis combined with R a m a n spectroscopic results, enables quantitative information on several components to be extracted e v e n w h e n one is d e a l i n g with c o m p l e x mixtures ~2-~5 In this paper, this method is used to analyze two different series of mixtures containing pure cholesterol and cholesterol esters (cholesterol oleate C18:1, linoleate C18:2, palmitate C16:0, and stearate C18:0) as a preliminary study These compounds were chosen because they are the predominant lipid species in the arterial intima The total amounts of C16:0, C18:0, C18:1, and C18:2 vary from 60 to 90% of the total cholesterol esters in the arterial wall Moreover, with increasing age and with atherosclerosis development, variations in the composition of cholesterol fatty acid esters appear There is a significant decrease in saturated fatty acids (palmitic and stearic) and an increase in linoleic acid and in the linoleate/oleate ratio 2,3.5,6 In the first step of this study, a R a m a n analysis of ternary mixtures of cholesterol C18:l, and C18:2 was used in order to examine the possibility of quantitatively determining the concentration of one particular c o m p o nent in the presence of the other two Next, quantitative analyses of binary mixtures of C16:0 and C18:0 were examined Both these cholesterol esters have a saturated fatty acid chain of 16 and 18 carbons, respectively Their R a m a n spectra therefore show practically no differences This second series of samples represents a real challenge as it is one of the most difficult analytical problems encountered MATERIAL AND METHODS Experimental Procedure Cholesterol, C18:1, and C18:2 were purchased from Sigma Chemical Company, and their purities were 95, 98, and 98%, respectively 0003-7028/96/5010-125352,00/0 © 1996 Society :tbr Applied Spectroscopy APPLIED SPECTROSCOPY 1253 73 72 71 % 70 I0 68 ~ 67 ~ / ~ M ~ ' / ~ / ~ / ~ X / / ~ 65// 12 13 ~ 15 ~%iX-~"~,~'X X , £ ~ , X X'~ 34 33 32 31 30 29 28 27 26 25 24 23 22 21 % CHOLESTEROL LINOLEATE Fro Ternary diagram representing the 65 calibration mixtures containing cholesterol, cholesterol linoleate, and cholesterol oleate C 16:0 and C 18:0 were purchased from Aldrich Company, and their purities were 97 and 96%, respectively Two sample series were prepared by using cholesterol and cholesterol ester solutions in chloroform mixed to obtain homogeneous mixtures These solutions were kept at r o o m temperature and in a vacuum until the total evaporation of chloroform occurred In the first series, 65 samples were prepared with molar fractions ranging from 63 to 72% for cholesterol, from to 14% for C18:1, and from 21 to 29% for C18:2, as shown in Fig Because concentrations of lipid components vary largely among atherosclerotic plaques, this range of concentrations was deliberately chosen as a first step The series of second samples was comprised of 49 mixtures of C16:0 and C18: with molar fractions ranging from 30 to 70% FT-Raman spectra were measured from 100 to 4000 cm -~ with a Bruker IFS 66 Fourier transform spectrometer equipped with a F R A 106 R a m a n scattering module The excitation source was a Spectra Physics T F R laser using 1047-nm radiation, and the detector used was a liquid nitrogen-cooled germanium diode Two successive spectra were obtained from each sample with 800-roW laser power in a 0 - ~ m - d i a m e t e r spot on the sample The spectra were collected with 200 scans at 2-cm-l resolution ( ~ total collection time) Data were recorded and processed with Bruker Opus software on a C o m p a q 386DX microcomputer The measured spectra were then transferred to a PC/AT 486 microcomputer equipped with a Turbo-Pascal PLS program Principle of Quantitative Raman Spectra Analysis Using the Partial Least-Squares Regression Quantitative R a m a n analysis with partial least-squares regression 15-21 was performed The R a m a n spectrum of a mixture can be represented as a vector, s = (I~ I~ I,,), in which /~ is the R a m a n intensity at wavelength i The sample composition is represented as another vector, c = (Ci Cj Cm), in which the C: are the concentrations of the mixture components The aim of quantitative R a m a n spectral analysis is the deduction of the composition from the sample Raman spectrum For this purpose, a model should be established during the calibration phase In practice, the sample composition is not completely 1254 Volume 50, Number 10, 1996 known in the case of complex mixtures, and the spectrum vector contains irrelevant information such as noise and variations in intensity due to unknown components and experimental variations This irrelevant information should be discarded during model elaboration The partial least-squares regression consists in calculating new spectral and concentration data by projecting the observed S and C data matrices of calibration samples onto the " f a c t o r s " bases, W and Q, respectively They are also the eigenvectors matrices of cov(tC.S) and cov(tS.C), respectively Under such conditions, the partial least-squares method takes the " a x e s " corresponding to the higher correlated variations between spectral information and known concentrations into account for the model elaboration The number of factors represents the amount of information involved and is therefore very significant for the estimation quality In practice, there are an optimal number of factors to be used, which should be determined by validating the model The relevant information is extracted and the irrelevant is discarded Moreover, with this method, each c o m p o n e n t of a mixture can theoretically be studied without any reference to the others Two approaches are then conceivable: either " P L S " , which consists in the determination of several components at the same time, or " P L S 1", in which each component is studied separately For each sample, the first R a m a n spectrum was used in the calibration phase and the second in the validating prediction phase In the prediction phase, the quality of the results was determined by calculating the standard error of prediction (SEP), defined as l SEP = Z IC,.,ro - C / uol2 i= /2 S where ns is the number of prediction samples RESULTS AND DISCUSSION Quantitative Determination of Cholesterol, C18:1, and C18:2 on Synthetic Mixtures Three spectra of ternary mixtures with their respective composition in cholesterol, C18:2, and C : I - - A (71.6, 21.6, and 6.8%); B (63.5, 29.2, and 7.3%), and C (63.2, 22.6, and % ) are shown in Fig 2, together with the spectra of the pure esters for comparison (Fig 3) There are only a few differences between these spectra In fact, the major vibrational bands observed in these spectra can be assigned to cholesterol vibrations (1667 c m - ~is due to the cholesterol C = C stretching vibration; 1455 and 1437 cm to the C H bending vibrations; 959, 880, 840, 805, 699, and 605 cm ~ to the vibrational modes of the sterol rings), and they are therefore shared by the three compounds Furthermore, R a m a n peaks due to fatty acid chains are often weak or overlapped with cholesterol peaks The main difference appears on the bands at 1667 and 1655 cm ~, assigned, respectively, to the stretching vibration of C = C bonds in cholesterol and in unsaturated fatty acids The relative intensity of the C = C band at 1655 cm ~ increases with the number of unsaturated double bonds per fatty acid chain However, these variations not allow a satisfactory quantitative analysis with the use of the stan- ' I ' tb- I CO ' ' ' I ' i ' ' m 1800 o o 1600 1400 1200 ! 000 Wavenumber 800 600 400 (cm -I ) FIG Near-infrared Raman spectra of pure cholesterol esters: (A) cholesterol oleate; (B) cholesterol linoleate; (C) cholesterol palmitate; (D) cholesterol stearate 1800 1600 1400 1200 i000 W a v e n u m b e r (cm 800 600 -i FIG Near-infrared Raman spectra of three ternary mixtures with respect to amounts of cholesterol, cholesterol linoleate, and cholesterol oleate: A (71.6, 21.6, 6.8%); B (63.5, 29.2, 7.3%); and C (63.2, 22.6, 14.2%) dard peak height method in the case of ternary mixtures With a PLS program, the best results were obtained by using the spectral region from 600 to 1800 cm -], which contains most bands of the three compounds The PLS method was carried out with the use of two different approaches to the problem For the first approach (PLS2), all concentration data were used at the same time in the model elaboration Under these conditions, the PLS method underscores the higher correlated variations between spectral information and all the concentrations Here, PLS only eliminates irrelevant information corresponding to experimental variations Free cholesterol and C18:2 molar fractions were used to calculate the PLS2 model (C18:1 concentration is deduced from the other two) For the second approach (PLS 1), only one concentration was used in each model elaboration With this method, PLS underscores the higher variations attributable to a specific compound and eliminates irrelevant information due to experimental variations and to the other compounds Three models corresponding to each component were run and compared to the PLS2 model Figures and show the variations of standard error of prediction vs number of factors used in the model elaborations for PLS2 and PLS1 methods, respectively As shown in Fig 4, standard errors of prediction decrease with the number of factors (NF) used in the calibration step until NF = 11 At that point, SEP values show no significant variations It should be noted that the model using only one factor provides a fairly good estimate of sample composition (SEP < 2%), which means that a large part of the relevant information is included in the first "dominant" factor No factors higher than 11 corre- sponding to irrelevant information improve model quality Therefore, the best PLS2 results were obtained with 11 factors in the model elaboration, and the standard errors of prediction calculated with 66 samples were 0.95, 1.19, and 1.30% for C 18:2, C 18:1, and free cholesterol, respectively In the case of P L S I , three models were independently calculated with the use of each of the three component concentrations Curves showing variations of SEP vs the -(3 Cholesterol ~ 1.8 Lio~e°~eate 1.6 r~ 1.4 1.2 0+8 I I I I I I I Number I I t I 12 of factors I I I I 15 I 18 I I 21 used FIG Plot of standard error of prediction (SEP) vs number of factors used in the calibration step for global quantitative analysis (PLS2) of the 65 ternary mixtures: ( ) for cholesterol concentration, (A) for cholesterol linoleate, and ( ) for cholesterol oleate APPLIED SPECTROSCOPY 1255 2.2 (3 - Cholesterol A Lin01eate • Oleate ',.0 k.O c-4 ,-~ ~ u3 o3 , cJ5 , I 1.8 1.6 1.4 1.2 0.8 I I I i I I I I I I I I t I 12 I 15 I I I 18 I I 21 1800 1600 1400 Wavenumber Number of factors used FIG used in (PLS1) (A) for Plot of standard error of prediction (SEP) vs n u m b e r of factors the calibration step for three independent quantitative analysis of the 65 ternary mixtures: (O) for cholesterol concentration, cholesterol linoleate, and ( • ) for cholesterol oleate number of factors used (Fig 5) were very similar to those obtained with PLS2 The best standard errors of prediction were 0.94% with the use of factors for cholesterol linoleate, 1.17% with 10-11 factors for cholesterol oleate, and 1.27% with 10 factors for free cholesterol The optimal number of factors used was not the same as in the PLS2 method nor the same for each compound This observation is consistent with the fact that information corresponding to the other component variations should be eliminated The results provided by PLS models were as good as those for the PLS2 This consideration means that NIR FT-Raman spectroscopy enables information to be obtained for one or several compounds in a multicomponent mixture This capability is of great interest in the case of arterial samples where only several compound compositions may be known or of interest The Quantitative Determination of C16:0 and C18: in Binary Synthetic Mixtures These two compounds were chosen because of the difficult differentiation of their Raman spectra Figure shows the spectra of C18: (A), C16:0 (C), and a mixture containing 50% of each (B) There is no visible difference between these spectra, because most bands observed on the spectra belong to groups that both compounds have in common Moreover, the bands assigned to the saturated fatty acid chain (1296 cm ~ due to C - H bending vibrations; 1129 and 1085 cm -~ due to C - C stretching vibrations) show very little variation in their relative intensity A quantitative analysis of C16:0 and C18:0 mixtures was carried out with the use of 49 samples for calibration and prediction The results are plotted in Fig The best standard error of prediction was 3.02% and was calculated with seven factors Moreover, the very important 1256 Volume 50, Number 10, 1996 1200 i000 800 600 (cm -I) FIG Near-infrared R a m a n spectra of (A) cholesterol stearate, (B) 50:50 cholesterol palmitate/cholesterol stearate mixture, and (C) cholesterol palmitate decrease of SEP in this case (from 10.95 to 3.02%) meant that the first seven factors had the same significance in the model elaboration There were no " d o m i n a n t " factors in this case, and the first seven factors had practically the same relevance This result confirms the fact that there are very few differences between the compounds I1 8.75 ~.' 6.5 4.25 I $ |1 13 15 t7 19 Number of factors used for calibration FIG '7 Plot of standard error of prediction (SEP) vs number of factors used in the calibration step for the quantitative analysis of 49 cholesterol palmitate/cholesterol stearate mixtures with concentrations ranging from 30 to 70% cholesterol palmitate The best standard error of prediction is 3.02% calculated with seven factors ' ' ' I ' ' ' [ ' '' I''' l''' I ' ' ' O3 ,-4 ~q £v A / cO CONCLUSION , 1800 1600 O~ oh tive analysis of cholesterol and cholesterol esters in the arterial wall Moreover, previous studies demonstrated that NIR F T - R a m a n could provide in situ information concerning other species that accumulate in the atherosclerotic arterial wall, such as elastin, collagen, carotenoids, and calcium apatite deposits 1~,22 Therefore, this technique should enable a quantitative study of disease progression and response to different modes of treatment to be made I 1400 , , l '1200 Wavenumber , , l i00'0 , , I 800 , , , 600 (cm -I) FIG Near-infrared Raman spectra of (A) an atherosclerotic rabbit aorta and (B) a ternary mixture containing cholesterol, cholesterol linoleate, and oleate Comparison of the Synthetic Mixture Spectra and Real Aorta S a m p l e s The results given in this paper demonstrate that N I R F T - R a m a n interfaced with the PLS methods can provide significant quantitative information on the composition of lipid mixtures Moreover, NIR F T - R a m a n can differentiate and analyze very similar compounds such as C16:0 and C18:0 In addition, this technique can determine biochemical composition of a single c o m p o u n d in a complex mixture This capability is of great importance in the case of biological samples, especially those from the arterial wall, in which m a n y species are present In practice, only several compounds are worth studying, because of their involvement in the pathogenesis Figure shows a comparison between a ternary mixture spectrum (cholesterol, C18:1, and C18: 2) and a thoracic aorta spectrum of a cholesterol-fed rabbit obtained in the following way: Prior to spectroscopic study, samples were homogenized in order to obtain spectroscopic results that could be c o m p a r e d to the biochemical ones Then, they were passively warmed to room temperature while being kept moist with isotonic saline solution The tissue samples were placed in an aluminium cuvette with a small amount of isotonic saline solution, to keep the tissue moist, and covered with a 0.1-ram glass window During the spectral measurements, the samples were cooled down with a cold nitrogen stream in order to keep them at the room temperature The spectra of the tissue samples were measured by using the same procedure used for the synthetic samples These two spectra have m a n y R a m a n bands in c o m m o n In particular, several bands of the thoracic aorta spectrum are significant (1667, 1655, 1455, 1437, 1301, and 699 cm-t) They are due to cholesterol and cholesterol esters These bands are easily discerned and can be used for an in situ quantita- The results to be found in this paper demonstrate that N I R F T - R a m a n spectroscopy interfaced with PLS regression can precisely determine the proportions of different fatty acids present in the arterial wall The method is a very powerful tool for histochemical analysis of atherosclerosis The next step of this study, in progress at the present time, is an in situ quantitative analysis of free and esterified cholesterol in cholesterol-fed rabbits arteries At the same time, the use of this technique will be extended to other lipidic and nonlipidic components that are involved in the pathogenesis of the human disease Once the method is validated, its extension to in situ human plaque studies will be envisaged O W Portman, Adv Lipid Res 8, 41 (1970) G T Malcom, J R Strong, and C Restrepo, Lab Invest 50, 79 (1984) D M Small, Arteriosclerosis 8, 103 (1988) B Lundberg, Atherosclerosis 56, 93 (1985) M E Rosenfeld, A Chait, E L Bierman, W King, P Goodwin, C E Walden, and R Ross, Arteriosclerosis 8, 338 (1988) J H Rapp, W E Connor, D S Lin, T Inahara, and J M Porter, J Lipid Res 24, 1329 (1983) Nguyen Quy Dao, M Jouan, R Plaza, and J Robieux, Brevet Franqais 84.15047 (1984) C Gomy, M Jouan, and Nguyen Quy Dao, Anal Chim Acta 215, l l (1988) D Phat, R Plaza, Nguyen Quy Dao, R N Vuong, and P G Steg, in Proceedings of the International Conference of Raman Spectroscopy, ICORS XII, J R Durig and J E Sullivan, Eds (John Wiley and Sons, New York, 1990), p 704 10 H Nguyen Thi Diem, D Phat, R Plaza, M Daudon, and Nguyen Quy Dao, in Proceedings of the Future Trends in Biochemical Applications of Lasers, SPIE Vol 1525 (SPIE, Bellingham, Washington, 1991), p 132 11 D Phat, P N Vuong, P Plaza, E Cheilan, and Nguyen Quy Dao, in Proceedings of the Future Trends in Biochemical Applications of Lasers, SPIE Vol 1525 (SPIE, Bellingham, Washington, 1991), pp 196-205 12 H Nguyen Quang, B Vu Thi, M Jouan, and Nguyen Quy Dao, Analusis 20~ 141 (1992) 13 H Nguyen Quang, M Jouan, and Nguyen Quy Dao, Appl Spectrosc 47, 2013 (1993) 14 M B Seasholtz, D D Archibald, A Lorber, and B R Kowalski, Appl Spectrosc 43, 1067 (1989) i5 T J Vickers and C K Mann, in Analytical Raman Spectroscopy, J G Grasselli and B J Bulkin, Eds (John Wiley and Sons, New York, 1991), pp 107-133 16 H Wold, in System Under Indirect Observations, K G J6reskog and H Wold, Eds (Elsevier-North Holland Publishing Company, Amsterdam, 1982), pp 1-54 17 M R Fuller, G L Ritter, and C S Draper, Appl Spectrosc 39, 73 (1988) 18 R Geladi and B R Kowalski, Anal China Acta 185, (1986) 19 A Lorbm; L E Wangen, and B R Kowalski, J Chemometrics 1, 19 (1987) 20 K R Beebe and B R Kowalski, Anal Chem 59, 1007 (1987) 21 D M Haaland and E V Thomas, Anal, Chem 60, 1193 (1988) 22 J J Baraga, M S Feld, and R R Rava, Proc Natl Acad Sci USA 89, 3473 (1992) APPLIED SPECTROSCOPY 1257 ... of factors used for calibration FIG ''7 Plot of standard error of prediction (SEP) vs number of factors used in the calibration step for the quantitative analysis of 49 cholesterol palmitate /cholesterol. .. Number of factors used FIG used in (PLS1) (A) for Plot of standard error of prediction (SEP) vs n u m b e r of factors the calibration step for three independent quantitative analysis of the... used FIG Plot of standard error of prediction (SEP) vs number of factors used in the calibration step for global quantitative analysis (PLS2) of the 65 ternary mixtures: ( ) for cholesterol concentration,