The inherent accuracy of 1H NMR spectroscopy to quantify plasma lipoproteins is subclass dependent

Atherosclerosis 190 (2007) 352–358

The inherent accuracy of 1H NMR spectroscopy to quantify plasma lipoproteins is subclass dependent Mika Ala-Korpela a,∗ , Niko Lankinen a , Aino Salminen a , Teemu Suna a , Pasi Soininen b , Reino Laatikainen b , Petri Ingman c , Matti Jauhiainen d , Marja-Riitta Taskinen e , K´aroly H´eberger f,∗∗ , Kimmo Kaski a a

Laboratory of Computational Engineering, Systems Biology and Bioinformation Technology, Helsinki University of Technology, Finland b Department of Chemistry, University of Kuopio, Finland c Department of Chemistry, Instrument Centre, University of Turku, Finland d Department of Molecular Medicine, Biomedicum, National Public Health Institute, Helsinki, Finland e Division of Cardiology, Department of Medicine, University of Helsinki, Finland f Chemical Research Center, Hungarian Academy of Sciences, Budapest, Hungary Received 24 February 2006; received in revised form 5 April 2006; accepted 7 April 2006 Available online 30 May 2006

Abstract Proton NMR spectroscopy as a means to quantify lipoprotein subclasses has received wide clinical interest. The experimental part is a fast routine procedure that contrasts favourably to other lipoprotein measurement protocols. The difficulties in using 1 H NMR, however, are in uncovering the subclass specific information from the overlapping data. The NMR-based quantification has been evaluated only in relation to biochemical measures, thereby leaving the inherent capability of NMR rather vague due to biological variation and diversity among the biochemical experiments. Here we will assess the use of 1 H NMR spectroscopy of plasma per se. This necessitates data for which the inherent parameters, namely the shapes and areas of the 1 H NMR signals of the subclasses are available. This was achieved through isolation and 1 H NMR experiments of 11 subclasses—VLDL1, VLDL2, IDL, LDL1, LDL2, LDL3, HDL2b , HDL2a , HDL3a , HDL3b and HDL3c —and the subsequent modelling of the spectra. The subclass models were used to simulate biochemically representative sets of spectra with known subclass concentrations. The spectral analyses revealed 10-fold differences in the quantification accuracy of different subclasses by 1 H NMR. This finding has critical significance since the usage of 1 H NMR methodology in the clinical arena is rapidly increasing. © 2006 Elsevier Ireland Ltd. All rights reserved. Keywords: 1 H NMR spectroscopy; Lipoprotein subclass; Quantification; Plasma; Serum; Coronary heart disease

1. Introduction In the 1 H NMR spectra of blood plasma one finds pronounced resonances originating from lipoprotein lipids [1]. ∗ Corresponding author at: Helsinki University of Technology, Laboratory of Computational Engineering, Systems Biology and Bioinformation Technology, P.O. Box 9203, FIN-02015 HUT, Finland. Tel.: +358 50 35 35 457; fax: +358 94 51 48 33. ∗∗ Corresponding author at: Chemical Research Center, Hungarian Academy of Sciences, H-1525 Budapest, P.O. Box 17, Hungary. Tel.: +36 1 438 1103x426. E-mail addresses: [email protected] (M. Ala-Korpela), [email protected] (K. H´eberger).

0021-9150/$ – see front matter © 2006 Elsevier Ireland Ltd. All rights reserved. doi:10.1016/j.atherosclerosis.2006.04.020

Two research groups independently developed a 1 H NMRbased lipoprotein quantification methodology in the mid 1990s [2,3] and several groups have been assessing this option later on as thoroughly discussed by Bathen et al. [4]. The majority of applications in which 1 H NMR spectroscopy has been used have focused on the main lipoprotein fractions, namely very low, low and high-density lipoproteins (VLDL, LDL and HDL, respectively) [1,3–12]. This is logical since the clinical recognition of lipoprotein subclasses in the coronary heart disease (CHD) risk assessment is quite recent [13]. The isolation of lipoproteins, needed to set up the NMRbased methodologies, is also tedious already at the level of the main fractions, but particularly if the distinct subclasses

M. Ala-Korpela et al. / Atherosclerosis 190 (2007) 352–358

are to be individually separated [3,14–16]. The operational NMR method, however, could rely on a single spectrum of a plasma sample and subsequent mathematical analysis of the subclass specific information, all achievable only in a few minutes [1–4]. In general, the results on the quantification of the main lipoprotein fractions in plasma by 1 H NMR have been in favour of the methodology. The resulting correlations, regardless of the mathematical analysis method applied, have been ranging from good to excellent between the NMR and the independent biochemical measures of mainly VLDL triglycerides as well as LDL and HDL cholesterol [4]. Recently also diffusion edited 1 H NMR spectroscopy has been introduced and applied to quantify lipoproteins from plasma, including subclasses [17–19]. However, it is only the work by Otvos et al. [2,20,21] which has directly included the lipoprotein subclasses as an integral part of the 1 H NMR protocol and developed it for the level of a commercial assay, named NMR LipoProfile® by LipoScience Inc. The NMR LipoProfile® test provides information on 15 lipoprotein subclasses [20]. Therefore, it is not surprising that the methodology is becoming increasingly used in clinical studies [22 and refs therein]. Recently considerable deviations were noticed between the NMR and gradient gel electrophoresis-based LDL particle size, somewhat questioning the comparability and clinical value of the NMR measures [23]. The recent advance in the understanding of atherothrombosis points to the clinical importance of lipoprotein subclasses in the CHD risk assessment [13] and so the existence of a practical method, such as 1 H NMR, for their quantification would be fundamental for clinical research. Our previous work supports the clinical usage of 1 H NMR through demonstrating that self-organising map analysis classifies plasma spectra in a clinically relevant manner at the level of the main lipoprotein fractions [7] and also at the level of lipoprotein subclasses [24]. In a very recent clinical application Festa et al. [22] present a modification of the previously described NMR LipoProfile® method. In this modification an improved spectral deconvolution is applied, which operates with more than 30 subpopulations when analysing the methyl region of the 1 H NMR spectra of plasma. However, the authors noted that the neighbouring NMR signals of each subpopulation differ only slightly in frequency and lineshape, and thereby the measurement reproducibility of the individual signal amplitudes is inherently limited. To overcome this limitation, the authors grouped neighbouring subpopulations empirically into a smaller number of subclasses so that the summed amplitudes of the individual subpopulation signals gave acceptable measurement precision (coefficient of variation <10%). By doing so Festa et al. [22] ended up investigating nine lipoprotein subclasses, namely large VLDL (including chylomicrons if present), medium VLDL, small VLDL, LDL, large LDL, small LDL, large HDL, medium HDL and small HDL. This reported change in the decon1H

353

volution method that has reduced the number of identifiable lipoprotein subclasses from the originally introduced 15 [20] to 9 [22] is somewhat confusing and calls for clarifying further work in figuring out the capabilities of 1 H NMR in studying lipoprotein subclasses. Therefore, in this paper we will present a determination of the inherent accuracy of 1 H NMR to quantify different lipoprotein subclasses.

2. Methods 2.1. Isolation of lipoprotein subclasses Eleven separate lipoprotein subclasses—VLDL1, VLDL2, IDL, LDL1, LDL2, LDL3, HDL2b , HDL2a , HDL3a , HDL3b and HDL3c —were isolated applying ultracentrifugation procedures as described previously [14–16]. For the 1 H NMR experiments each of the isolated lipoprotein subclass sample was dialysed overnight against a buffer containing 137 mM NaCl, 2.7 mM KC1, 10 ␮M Na2 EDTA, 10 mM Na-phosphate, pD 7.0 in D2 O. 2.2. 1 H NMR spectroscopy The 1 H NMR spectra of the 11-lipoprotein subclass samples were recorded at the physiological temperature of 310 K on a Bruker AVANCE 600 spectrometer operating at 600.13 MHz. Details of the NMR experiments are presented in supplementary methods. 2.3. Modelling and simulation of spectra The description and rationale of the simulated data set used here can be found in supplementary methods; it relates to our recent work on the classification of lipoprotein subclass profiles on the basis of 1 H NMR spectra [24]. Briefly, two biochemically distinct categories of spectra with markedly different lipoprotein subclass profiles corresponding to ‘normal’ (marked N) and ‘metabolic syndrome’ (marked MS) were generated. In addition, a category of spectra representing a metabolic pathway (marked MP) between the N and MS categories was generated. The simulated spectra containing the lipoprotein subclass part of actual plasma are called spectra of lipoplasma [24]. As illustrated in Fig. 1, the spectra of lipoplasma were constructed using the 11 lipoprotein subclass model signals obtained from the lineshape fitting analyses as explained in supplementary methods (see also Supplementary Table 1 and Supplementary Fig. 1). In total, a set of 729 spectra were simulated for both the N and MS categories and 501 spectra for the MP category. Separate random categories of 501 and 250 spectra, in which the areas of each lipoprotein model signal group were randomly chosen within the minimum and maximum of the N, MS and MP categories, were simulated to allow for more extensive variation in the training and independent testing of the lipoprotein subclass quantifications. Random experimen-

354

M. Ala-Korpela et al. / Atherosclerosis 190 (2007) 352–358

2.4. Quantification of lipoprotein subclasses Partial least squares (PLS) is a widely used and accepted chemometric method in various disciplines [25,26], including NMR spectroscopy [27]. It is a method for relating the variations in one or several response variables (Y-variables or dependent variables) to the variations of several predictors (X-variables) with explanatory or predictive purposes [25,26]. PLS performs particularly well when the various Xvariables express common information, i.e., when there is a large amount of correlation, or even collinearity among them. This condition applies very well in this particular case of 1 H NMR spectroscopy data. The terminal CH3 region in the lipoplasma spectra were analysed using 1000 spectral data points. A separate quantitative model was constructed for each lipoprotein subclass resulting in 11 PLS models for each three-noise levels. The training data set used here consisted of 1094 spectra in the N and MS categories and of 751 spectra in the MP and random categories, i.e., in total of 1845 spectra. All PLS models were tested with the 250 random spectra; all the results presented in this communication are from these test runs in which only randomly generated spectra, unforeseen by the PLS models, were used. All the analyses were performed on a laptop PC with a Pentium III, 2.0 GHz processor. The PLS analyses took a few minutes.

3. Results and discussion

Fig. 1. Illustration of the 1 H NMR spectral lineshapes for the 11 lipoprotein subclasses VLDL1, VLDL2, IDL, LDL1, LDL2, LDL3, HDL2b , HDL2a , HDL3a , HDL3b and HDL3c and for some lipoplasma spectra in the terminal methyl CH3 region (around 0.85 ppm) indicated by the reddish bars. At the bottom a spectrum representing a normal lipoprotein subclass profile is shown together with key resonance assignments. The lineshapes shown for the lipoprotein subclasses are from lineshape fitting analyses of the experimental 1 H NMR spectra measured at 310 K. The spectra termed lipoplasma are varied sums of these subclass model signals; N refers to spectra representing normal lipoprotein subclass profiles, MS refers to spectra representing subclass profiles typical for a metabolic syndrome, and MP refers to spectra representing a metabolic pathway between N and MS. The lipoplasma spectra termed RANDOM are from the category of 250 spectra, in which the areas of each lipoprotein model signal were randomly chosen within the minimum and maximum of the N, MS and MP categories, simulated to allow for an independent testing of the lipoprotein subclass quantifications.

tal noise was added to each simulated lipoplasma spectrum at the levels of 0.2, 1.0 and 5.0% of the highest intensity of the basis spectrum in the N category; the lowest noise levels roughly correspond to typical set-ups, which are quite easily attainable in these kinds of experiments [3].

Fig. 1 illustrates the characteristics of the 1 H NMR spectra for the isolated lipoprotein subclasses and for the simulated lipoplasmas. Here we have focused on the methyl resonance region since it is the most commonly used signal area in relation to lipoprotein analysis by 1 H NMR and also used in the commercial NMR LipoProfile® test [1–4,20,21]. The CH3 resonances show remarkable similarity for all the lipoprotein subclasses with an intense distorted triplet like signal at around 0.85 ppm and some cholesterol backbone related resonances. The size dependent chemical shift of all the resonances in the CH3 region is evident [28]. In the CH3 resonances of the lipoplasma spectra one can see a clear difference between the MS and N categories. The relative increase of triglyceride rich VLDLs and IDL in the MS spectrum is also quite clearly noticeable but other differences between the MS and N spectra, i.e., decreased LDL1, increased LDL2 and LDL3, decreased HDL2 ’s and increased HDL3 ’s for the MS spectrum, are much more difficult to pick up by eye (see Supplementary Table 1 for the details of the MS and N categories). In the light of this visual consideration, it may be surprising that deconvolution of only the main CH3 resonance is able to quantify 15 lipoprotein subclasses [20]. In Table 1 the results of the quantitative PLS test analyses of the CH3 regions of the 250 random lipoplasma spectra at all noise levels are given. Some correlations between the known lipoprotein subclass areas and the ones obtained by

M. Ala-Korpela et al. / Atherosclerosis 190 (2007) 352–358

355

Table 1 The R2 values and the average relative errors between the known simulated and the 1 H NMR and PLS predicted lipoprotein subclass areas in the random test set of 250 lipoplasma spectra

a Random experimental noise was added to each simulated spectrum at the level of 0.2, 1.0 or 5.0%

of the highest intensity of the basis spectrum in the N category. average relative error in percentage (± standard deviation) calculated for each subclass (and

b The

250

((|APLS −Aknown |/Aknown )×100%)

i i i i=1 , in which APLS denotes the for all the noise levels) as ARE = i 250 1 H NMR signal area of the lipoprotein subclass in question obtained via the PLS analysis, Aknown i is the corresponding simulated 1 H NMR signal area and i goes through all the spectra in the random test set.

the quantitative PLS analyses of the CH3 regions of the 250 random spectra are illustrated in Fig. 2 for the noise level of 1.0% together with the average relative errors (±standard deviations) of the 1 H NMR + PLS quantifications for each

lipoprotein subclass. The results given in Table 1 show that at the noise levels of 0.2 and 1.0% the R2 ’s range from 0.9816 to 1.0000 and the average relative errors from 0.09 to 4.19%. This indicates that the PLS analyses, for the completely inde-

Fig. 2. Illustration of some representative correlations between the known lipoprotein subclass areas (simu) and the ones obtained by the quantitative PLS analysis of the CH3 regions of the 250 random lipoplasma spectra (1 H NMR + PLS) at the noise level of 1.0%. Each dot in the correlation figures represents one lipoplasma spectrum. The lines drawn and the R2 values are from the linear regressions for the data (see also Table 1). The plot on the right shows the average relative error (± standard deviation) of the 1 H NMR + PLS quantifications for each lipoprotein subclass. Note the clear differences between the subclasses and the considerably decreased accuracy particularly in the case of LDLl quantification.

356

M. Ala-Korpela et al. / Atherosclerosis 190 (2007) 352–358

pendent random test set of the lipoplasma spectra, are able to reach good quantification for each lipoprotein subclass. In relation to these results, it should be noted that noise levels less than 1.0% are experimentally quite routinely attainable [1,3]. As indicated in Fig. 2, the average relative errors of the quantifications denote noteworthy differences in the capability of the (PLS analyses of the) 1 H NMR spectra to uncover the subclass specific signal areas. The VLDL and HDL3 subclasses give the most accurate results. The accuracy seems to be considerably less for the HDL2 subclasses as well as for the IDL and for the smaller LDL’s. At all noise levels there is approximately a 10-fold difference in the quantification accuracy between the most accurate (VLDLs and HDL3 ’s) and the least accurate (LDL1) values. It should be noted that in a real situation 1 H NMR signals from serum albumin and albumin bound fatty acids cause an overall broad hump at the aliphatic region of the spectra overlapping the resonances of lipoprotein lipids (see Supplementary Fig. 1). Thus, the results reported here represent an optimum case obtainable by 1 H NMR for lipoprotein subclass quantification; it is likely that the relative accuracies between the lipoprotein subclasses transform for the real cases but the absolute errors are noticeably higher for real samples. As few as 6 PLS components (at each noise level) were needed in the PLS model to approach a near plateau value of the predicted residual sum of squares. Including further PLS components decreased the prediction error only marginally. In fact, we also performed quantitative artificial neural network (ANN) analysis [6] for exactly the same spectra as used in the PLS modelling. These results (data not shown) for the subclass quantification were almost identical to those presented in Table 1 as obtained by PLS. Thus, the consistent results from the PLS (and ANN) analyses support the conclusion that these quantification results reflect the inherent accuracy of the 1 H NMR data in the case of lipoprotein subclass quantification. The general trends in the results for the quantification accuracy of the different lipoprotein subclasses presented in Fig. 2 become to a large extent understandable by looking at the spectral information shown in Fig. 1. The most accurate results are obtained for the VLDL and HDL3 subclasses that are on the high and low frequency sides, respectively, of the methyl region. The intrinsic, lipoprotein particle size dependent chemical shift behaviour favours the resolution of the smaller particles, i.e., HDL3 ’s [28]. For the VLDL’s the methyl signals appear sharper than those of the other subclasses and also quite clearly shifted towards higher frequencies. Even though the size dependent chemical shift for the larger particles is not as pronounced as for the smaller particles [28], there are, however, much bigger size differences in the case of VLDL particles in comparison to IDL and LDL particles; the VLDL particles range typically from 30 to 80 nm while the IDL and LDL particles are from 16 to 28 nm in diameter. The methyl resonances of the IDL and HDL2 subclasses fall in between the VLDL and HDL3 res-

onances, and the LDL resonances are in the middle of the methyl signal of the spectra. The similarity of the methyl signals, particularly for the LDL and HDL2 subclasses, and the small size differences within the LDL subclasses, i.e., only slightly differing chemical shifts of the methyl resonances of the LDL subclasses, make the decreased quantification accuracy for the IDL, LDL and HDL2 subclasses understandable. Why the accuracy for the LDL2 and LDL3 quantification appears to be three times better than that for the LDL1 seems visually unexplainable, but remains an empirical fact. Quantification of lipoprotein lipids, whether in main fractions or in subclasses, using 1 H NMR spectroscopy and a model-based approach [1–3,20] relies on a biophysical assumption that the 1 H NMR lineshapes of the lipoprotein particles are identical within each lipoprotein subclass. This means that the molecular composition and structure, affecting the NMR visibility of the lipids in the lipoprotein particles, is assumed to be alike for all individuals. However, it is known that there is individual variation in the chemical composition, and thus also in the 1 H NMR signals of the lipoprotein particles [1]. This biological variation will obviously always be present in a real data set (whether an explicit model would be constructed for the analyses or not). Another assumption in using 1 H NMR to quantify lipoproteins is that the mathematical analysis used is capable of revealing the (sub)class specific information from the heavily overlapping data. When studying real plasma samples, and comparing the results based on mathematical analyses of the experimental 1 H NMR data with independent biochemical lipid measurements, both of the abovementioned assumptions are implicitly present and inseparable from each other. It seems that our study is the first in which these two assumptions have been separated and the intrinsic capability of 1 H NMR spectroscopy in quantifying lipoprotein subclasses is addressed in a set of simulated spectra with known amounts (i.e., known NMR signal areas) of the subclasses and thus without the effects of individual biological variation and unknown inaccuracies in the biochemical lipid analyses [4]. Naturally, simulation of experimental data always provides limited information as compared to real experiments. Nevertheless, here the simulated spectra represent an optimum case from the 1 H NMR methodology point of view, for example, a decrease in the magnetic field strength (i.e., decreased resolution) or an increased number of subclass species will intensify the results and trends obtained. The signal-to-noise ratio of the 1 H NMR data also appears crucial for the quantification accuracy of the lipoprotein subclasses. Bearing in mind the abovementioned assumptions and limitations we now briefly compare the present results with existing literature on lipoprotein subclass quantification by 1 H NMR [17–22]. As already indicated, the NMR LipoProfile® test has established a commercial stage but it seems that no results comparing the LipoProfile® values and other measures for the individual subclass quantifications have been published. It would be of great interest to know how the NMR LipoProfile® values compare to independent

M. Ala-Korpela et al. / Atherosclerosis 190 (2007) 352–358

biochemical measures of lipoprotein subclasses as well as to how the special deconvolution software, used as an integral part of the LipoProfile® test, would cope with the overlapping methyl resonances in the case of known signal areas, i.e., in a case as presented here using quantitative PLS analysis. Three papers appear in the literature in which comparisons with 1 H NMR-based lipoprotein subclass quantifications have been compared with independent biochemical lipid measurements [17–19]. In two of these studies [17,18], the NMR experimentation was performed at 303 K, at a temperature that the core cholesterol esters of LDL particles are likely to be partly in a liquid crystalline state thereby diminishing considerably their NMR visibility and also introducing additional variability due to individual differences in the lipid and fatty acid composition of the particles. However, in the very recent work of Dyrby et al. [19] the NMR experimentation was performed at 318 K, i.e., well above the phase transition temperature of the core cholesterol esters of LDL. This study is particularly important in providing the first NMR LipoProfile® independent information on 1 H NMR-based lipoprotein subclass analysis. Only seventeen volunteers, though with a broad range of lipid levels, were the basis of the study, thereby limiting any far reaching conclusions. The trend in the quantification errors (root mean squared errors of cross-validation) presented by Dyrby et al. [19], however, is quite similar to our results for the average relative errors. For example, they found that the quantification error was the smallest for the smallest HDL and that the middle LDL subclasses (termed LDL2-5 in [19]) had approximately four times larger error than the two other LDL subclasses (termed LDL1 and LDL6 in [19]). These findings, also evoking the previous discussion in relation to the spectral information content and heavy overlap of the methyl resonances of LDL subclasses, suggest that the inherent capability of 1 H NMR to accurately quantify individual LDL subclasses is limited to a certain extent. It should be emphasised, that a key issue affecting the accuracy of various lipoprotein subclass values, is the number of subclasses quantified from the NMR data. The presented result is thus in close connection to the lipoprotein subclass isolation protocol, i.e., 11 subclasses, used. Another number of isolated lipoprotein subclasses would result in slightly different quantitative values; the more subclasses to be quantified, the more inaccuracies would be imposed, particularly in the case of the most overlapping subclasses in the 1 H NMR spectra, namely the LDL subclasses.

4. Conclusion The inherent capability of 1 H NMR to quantify lipoprotein subclasses depends on the particular subclass. Ten-fold differences in the quantification accuracy were observed between the most accurate VLDL as well as HDL3 subclasses and the least accurate LDL1 subclass. These finding are notewor-

357

thy in relation to the clinical utility of lipoprotein subclass quantification by 1 H NMR spectroscopy.

Acknowledgements M. A-K thanks Dr. Jukka Heikkonen for valuable discussions. This work has been supported by the Academy of Finland and partly by the Hungarian Scientific Foundation: OTKA T037684.

Appendix A. Supplementary data Supplementary data associated with this article can be found, in the online version, at doi:10.1016/j.atherosclerosis. 2006.04.020.

References [1] Ala-Korpela M. 1 H NMR spectroscopy of human blood plasma. Progr Nucl Magn Reson Spectr 1995;27:475–554. [2] Otvos JD, Jeyarajah EJ, Bennett DW, Krauss RM. Development of a proton nuclear magnetic resonance spectroscopic method for determining plasma lipoprotein concentrations and subspecies distributions from a single, rapid measurement. Clin Chem 1992;38:1632– 8. [3] Ala-Korpela M, Korhonen A, Keisala J, H¨orkk¨o S, Korpi P, Ingman LP, Jokisaari J, Savolainen MJ, Kes¨aniemi YA. 1 H NMR-based absolute quantitation of human lipoproteins and their lipid contents directly from plasma. J Lipid Res 1994;35:2292–304. [4] Bathen TF, Krane J, Engan T, Bjerve KS, Axelson D. Quantification of plasma lipids and apolipoproteins by use of proton NMR spectroscopy, multivariate and neural network analysis. NMR Biomed 2000;13:271–88. [5] Hiltunen Y, Heiniemi E, Ala-Korpela M. Lipoprotein-lipid quantification by neural-network analysis of 1 H NMR data from human blood plasma. J Magn Reson B 1995;106:191–4. [6] Ala-Korpela M, Hiltunen Y, Bell JD. Quantification of biomedical NMR data using artificial neural network analysis: lipoprotein lipid profiles from 1 H NMR data of human plasma. NMR Biomed 1995;8:235–44. [7] Kaartinen J, Hiltunen Y, Kovanen PT, Ala-Korpela M. Application of self-organizing maps for the detection and classification of human blood plasma lipoprotein lipid profiles on the basis of 1 H NMR spectroscopy data. NMR Biomed 1998;11:168–1676. [8] Bell JD, Sadler PJ, Macleod AF, Turner PR, La Ville A. 1 H NMR studies of human blood plasma. Assignment of resonances for lipoproteins. FEBS Lett 1987;219:239–43. [9] Otvos JD, Jeyarajah EJ, Hayes LW, Freedman DS, Janjan NA, Anderson T. Relationships between the proton nuclear magnetic resonance properties of plasma lipoproteins and cancer. Clin Chem 1991;37:369–76. [10] Engan T, Bjerve KS, Hoe AL, Krane J. Proton magnetic resonance spectroscopy of fractionated plasma lipoproteins and reconstituted plasma from healthy subjects and patients with cancer. Scand J Clin Lab Invest 1992;52:393–408. [11] Serrai H, Nadal L, Leray G, Leroy B, Delplanque B, de Certaines JD. Quantification of plasma lipoprotein fractions by wavelet transform time-domain data processing of the proton nuclear magnetic resonance methylene spectral region. NMR Biomed 1998;11:273–80. [12] Bathen TF, Engan T, Krane J. Principal component analysis of proton nuclear magnetic resonance spectra of lipoprotein fractions from

358

[13] [14]

[15]

[16]

[17]

[18]

[19]

M. Ala-Korpela et al. / Atherosclerosis 190 (2007) 352–358 patients with coronary heart disease and healthy subjects. Scand J Clin Lab Invest 1999;59:349–60. Superko HR. Lipoprotein subclasses and atherosclerosis. Front Biosci 2001;6:355–65. Taskinen MR, Packard CJ, Shepherd J. Effect of insulin therapy on metabolic fate of apolipoprotein B-containing lipoproteins in NIDDM. Diabetes 1990;39:1017–27. Griffin BA, Caslake MJ, Yip B, Tait GW, Packard CJ, Shepherd J. Rapid isolation of low density lipoprotein (LDL) subfractions from plasma by density gradient ultracentrifugation. Atherosclerosis 1990;83:59–67. Botnia Study Group. Tilly-Kiesi M, Knudsen P, Groop L, Taskinen MR. Hyperinsulinemia and insulin resistance are associated with multiple abnormalities of lipoprotein subclasses in glucose-tolerant relatives of NIDDM patients. J Lipid Res 1996;37:1569–78. Dyrby M, Petersen M, Engelsen SB, Nørgaard L, Sidelmann UG. Quantification of lipoprotein subtractions using 1 H-NMR and chemometrics, in Magnetic resonance in food science. In: Webb GA, Belton PS, Gil AM, Rutledge DN, editors. Latest developments. Cambridge: The Royal Society of Chemistry; 2002. p. 101–9. Petersen M, Dyrby M, Toubro S, Engelsen SB, Norgaard L, Pedersen HT, Dyerberg J. Quantification of lipoprotein subclasses by proton nuclear magnetic resonance-based partial least-squares regression models. Clin Chem 2005;51:1457–61. Dyrby M, Petersen M, Whittaker AK, Lambert L, Nørgaard L, Bro R, Engelsen SB. Analysis of lipoproteins using 2D diffusion-edited NMR spectroscopy and multi-way chemometrics. Anal Chim Acta 2005;531:209–16.

[20] Otvos JD. Measurement of lipoprotein subclass profiles by nuclear magnetic resonance spectroscopy. Clin Lab 2002;48:171–80. [21] Otvos J, Jeyarajah E, Bennett D. A spectroscopic approach to lipoprotein subclass analysis. J Clin Lig Assay 1996;19:184–9. [22] Festa A, Williams K, Hanley AJ, Otvos JD, Goff DC, Wagenknecht LE, Haffner SM. Nuclear magnetic resonance lipoprotein abnormalities in prediabetic subjects in the Insulin Resistance Atherosclerosis Study. Circulation 2005;111:3465–72. [23] Witte DR, Taskinen MR, Perttunen-Nio H, Van Tol A, Livingstone S, Colhoun HM. Study of agreement between LDL size as measured by nuclear magnetic resonance and gradient gel electrophoresis. J Lipid Res 2004;45:1069–76. [24] Suna T, Salminen A, Soininen P, et al. 1 H NMR metabonomics of plasma lipoprotein subclasses—an elucidation of metabolic clustering by self-organising maps. NMR Biomed, submitted for publication. [25] Geladi P, Kowalski B. Partial least squares regression: a tutorial. Anal Chim Acta 1986;185:1–17. [26] Wold S. Chemometric methods for molecular design. In: van de Waterbemmd H, editor. PLS for multivariate linear modeling. Weinheim: VCH; 1995. p. 195–218 [Chapter 4.4]. [27] Holmes E, Antti H. Chemometric contributions to the evolution of metabonomics: mathematical solutions to characterising and interpreting complex biological NMR spectra. Analyst 2002;127: 1549–57. [28] Lounila J, Ala-Korpela M, Jokisaari J, Savolainen MJ, Kes¨aniemi YA. Effects of orientational order and particle size on the NMR line positions of lipoproteins. Phys Rev Lett 1994;72:4049–52.