1288 Raquel Olias Barbara Maldonado Pauline Radreau Gwénaëlle Le Gall Francis Mulholland Ian J. Colquhoun E. Katherine Kemsley Institute of Food Research, Norwich, United Kingdom
Received June 28, 2005 Revised September 20, 2005 Accepted September 21, 2005
Electrophoresis 2006, 27, 1288–1299
Research Article
Sodium dodecyl sulphate-polyacrylamide gel electrophoresis of proteins in dry-cured hams: Data registration and multivariate analysis across multiple gels This study investigates whether dry-cured hams from two European countries can be distinguished using SDS-PAGE. Thirty-seven commercial hams (19 Spanish, 18 French) were used in the study. Four protein fractions were extracted from each sample, with sufficient material prepared to allow each fraction to be analysed in triplicate lanes. The complete extraction process was carried out in duplicate. The 24 specimens originating from each ham sample were randomly allocated to different lane positions and gels, as were at least two reference lanes (for reference proteins). In total, 118 gels were prepared. Mathematical routines were developed using a matrix language to process the gel image files. Procedures were written to carry out ‘withingel’ image correction, lane extraction and normalization, ‘between-gel’ data registration and linear discriminant analysis (LDA) of each fraction’s data to establish whether the provenance could be systematically distinguished. The between-gel registration was carried out using a genetic algorithm (GA). Feature selection was also performed using a GA, to pass subsets of features to the LDA routine. Cross-validated classification success rates were 84, 91, 81 and 85%, respectively, for the four fractions. We conclude that SDS-PAGE can be conducted in a sufficiently quantitative manner and can potentially verify the provenance of regional speciality dry-cured hams. Keywords: Discriminant analysis / Genetic algorithms / Ham / Multivariate analysis / SDS-PAGE DOI 10.1002/elps.200500466
1 Introduction Dry-cured ham is a traditional food product that is highly appreciated for its distinctive aroma, taste and texture. Among the best-known examples are Parma and San Daniele hams from Italy, Iberian and Serrano hams from Spain and Bayonne hams from France. The character of the ham is determined by a combination of the raw material (breed of pig, diet etc.) and the processing. Although the processing follows a common pattern with stages of salting, washing, salt equilibration and drying/ ripening, the conditions of temperature, humidity and Correspondence: Dr. E. Katherine Kemsley, Institute of Food Research, Norwich Research Park, Colney, Norwich, NR4 7UA, United Kingdom E-mail:
[email protected] Fax: 144-1603-507-723 Abbreviations: GA, genetic algorithm; LDA, linear discriminant analysis; PLS, partial least squares; PCA, principal component analysis
© 2006 WILEY-VCH Verlag GmbH & Co. KGaA, Weinheim
time applied at each stage vary according to the country and region [1]. A common feature of all these uncooked hams is the extensive breakdown of muscle proteins by endogenous enzymes that takes place during the lengthy maturation period (6–24 months depending on the type of ham). Proteolysis by cathepsins and calpains produces polypeptides that are partly degraded by peptidases to small peptides and free amino acids. The extent of proteolysis largely determines the texture of the final product, while the peptides and amino acids contribute directly to taste and, through further reactions, to the generation of aroma volatiles. Previous studies have established methods for the extraction and fractionation of proteins in drycured hams [2] and have used SDS-PAGE to determine how the profile changes with time in the course of processing the raw ham into the finished cured product [3, 4]. The aim of this study was to investigate whether cured hams from two different European countries of origin (France and Spain) could be systematically distinguished www.electrophoresis-journal.com
General
Electrophoresis 2006, 27, 1288–1299
using SDS-PAGE. In previous work, protein profiles from IEF have been analysed using multivariate methods to discriminate between wheat [5, 6] and potato [7] varieties and to determine the composition of meat mixtures containing different species [8] or mechanically recovered meat [9]. The novelty of the work in the present paper lies in the rigorous quantitative manner in which the data are analysed: there is little precedent in the literature of quantitative analysis of SDS-PAGE data on this scale, or using the combination of multivariate methods as presented here.
2 Materials and methods 2.1 Materials Commercial hams of verified provenance were selected by experts working in the European Union (EU) project TYPIC (www.typic.org) in order to be representative of the range of products available to consumers in the two countries. The project is studying differences between ‘typical’ and related ‘nontypical’ (e.g. branded) products as perceived by the consumer and as measured by various objective methods such as protein profiling. A quantitative analysis of the aroma volatiles in some of the same hams has been published [10]. Hams are included with differences in the breed, region of origin (Table 1), length of maturation, scale of production, and with or without a label such as ‘Protected Denomination of Origin’. Here we only focus on one potential classifier, the country of origin: 19 Spanish and 18 French samples were used in the study. Samples were supplied in vacuum packs of 100 g (about six slices in each) and stored at 47C until opening. Table 1. Details of French and Spanish dry-cured hams Country (no. samples)
Breed (no. samples)
Region (no. samples)
France (18)
White (16)
Auvergne (7) Lacaune or Aveyron (5) Bayonne (3) Miscellaneous (3)a)
Basque or Bigorre (2) Spain (19)
White (11) Iberian (8)
Teruel (6) Guijelo, Huelva or Cordoba (7) Miscellaneous (6)a)
a) Large-scale production including supermarket own brands. © 2006 WILEY-VCH Verlag GmbH & Co. KGaA, Weinheim
1289
2.2 Sample preparation Several slices were taken from an opened pack and the fat was cut from the ham. The remaining muscle tissue (all muscle types) was cut into small pieces in a mortar and frozen by pouring liquid nitrogen on top. The frozen pieces were broken up with a pestle. After the nitrogen had evaporated, they were placed on dry ice and ground further with a hand blender until a fine powder was obtained. The powder was stored at 2807C. The extraction was based on the method of Toldrá et al. [2] but carried out on a smaller scale. Also, four protein fractions were extracted from each sample rather than two: sufficient material was prepared to allow each fraction to be analysed in triplicate lanes. The complete extraction process was carried out in duplicate; all steps were performed at 47C (on ice). Powdered ham (100 mg) was placed in an Eppendorf tube and 1 mL of buffer A (30 mM potassium phosphate, pH 7.3) was added. The sample was vortexed (47C room) for 15 min and then centrifuged. The supernatant was taken (leaving the upper layer of fat), divided into four 200 mL portions and placed in small Eppendorfs (‘fraction 1’). Three of these were used for the gel lane repeats and the fourth to establish the protein concentration. The pellet from this separation was mixed with 1 mL of buffer B (30 mM potassium phosphate with 1.1 M KCl), vortexed for 15 min at 47C and centrifuged. The supernatant was taken and divided into four portions as before (‘fraction 2’). The pellet from this separation was re-extracted with the same solvent in a repeat of the above procedure (‘fraction 3’). The remaining pellet was extracted into 1 mL of buffer C (50 mM Tris (trizma), pH 7.5, 8 M urea, 2 M thiourea, 75 mM DTT, 3% SDS), vortexed 15 min at 47C, centrifuged and the supernatant (‘fraction 4’) again divided into four portions as before. Preliminary studies (not reported here) had shown that analysis of fraction 4 was more repeatable when the intermediate solubility proteins were extracted by the preceding two-step procedure as described, rather than by a single extraction into 1.1 M KCl.
2.3 Electrophoresis, gel staining and imaging The same quantity of protein (1 mg) was loaded onto each gel lane. Protein concentration was determined for each sample/fraction in the fourth tube by Bradford’s method [11]. Then the calculated volume (X mL) containing 1 mg of protein for each sample replicate was added to 2.5 mL NuPAGE® LDS sample buffer, 1.0 mL NuPAGE reducing agent and 102(X 1 3.5) mL of deionized water to give a total volume of 10 mL. The samples were heated at 707C for 10 min. SDS-PAGE was carried out using 4–12% www.electrophoresis-journal.com
1290
R. Olias et al.
Novex Bis-Tris gels (NuPAGE, Invitrogen) with the running buffer MOPS (Invitrogen). An experimental design was generated that randomly allocated the 24 specimens originating from each ham sample to different lane positions and gels. Also allocated to random positions on each gel were two reference lanes containing the molecular weight standard (precision plus protein™, BioRad, Hercules, USA), plus one or more control lanes containing 1 mg of BSA (1 mg/mL). In total, 118 gels were prepared, each including up to nine lanes of sample data. After electrophoresis the gels were washed in water for 20 min, incubated in SYPRO Ruby stain overnight with continuous gentle agitation and then washed in 10% methanol/6% acetic acid for 1 h. The ProXPRESS proteomic imaging system with ProFinder imaging software was used to image the gels. Top illumination was used with a 480/30 excitation filter and a 630/30 emission filter at 100 mm resolution. The resulting apf image files were converted to 16-bit tif image files and processed as described below.
Electrophoresis 2006, 27, 1288–1299
2.4.1.2 Identification of lanes and extraction of sample data vectors Following image correction, lanes of sample data were identified, again in a semiautomated manner, using the following procedure: the Fourier transform of each pixel column was calculated, and the non-DC components integrated; if this value (i.e. the integrated power spectrum) was greater than a threshold value for the gel, Then the pixel column was deemed to contain sufficient ‘information’ (i.e. multiple protein bands). This process was used to define a rectangle for each occupied lane in the image. Lanes corresponding to reference proteins were identified by referring to the experimental design. For each lane of sample data, the average of all pixel columns in the rectangle was calculated, to yield a single 431-element vector. These steps are illustrated in Fig. 1b.
2.4.1.3 Normalization 2.4 Data analysis All data analysis was carried out using Matlab (The Mathworks, Cambridge, UK). Analysis comprised two distinct stages: preprocessing and statistical modelling. The preprocessing stage was subdivided into two steps:
2.4.1 Preprocessing step 1: within-gel preprocessing 2.4.1.1 Correction of the 16-bit greyscale images for distortion, and simultaneous within-gel registration of lanes on the gel to selected reference protein bands Image correction is necessary to correct for barrel- and pincushion-type image distortion, which causes individual protein bands to deviate from the horizontal. An algorithm was written in house to carry this out. The correction procedure comprised piecewise linear interpolation of each column of greyscale pixel values, with abscissae ranges for each piece defined by third-degree polynomial functions of the pixel column index. (The polynomials were themselves obtained by fitting to user-defined control points, identified to coincide with four selected reference bands.) Correction was thus performed simultaneously with registration to four reference proteins. The regions of the gel image above and below the top and bottom polynomials were discarded; after interpolation, each column contained 431 pixels. These procedures are illustrated in Fig. 1a. © 2006 WILEY-VCH Verlag GmbH & Co. KGaA, Weinheim
The vectors of sample data were normalized by scaling, such that the minimum element in each was equal to zero, and the sum of elements equal to 1. This step is illustrated in Fig. 1c.
At this stage, any lanes that were clearly faulty (for example, containing errors in the images, such as saturation) were discarded. In total, 862 data vectors were retained, giving a complete data matrix of dimensions (8626431). Principal component analysis (PCA) (see Section 2.4.3 and Section 3) applied to this matrix showed that its variance (information content) is dominated by differences between fractions. Visual examination of the data suggested that registration to correct for misalignment of bands between gels would be helpful. It was therefore decided to further preprocess the data with a second registration step and, furthermore, to do this for the data from each protein fraction separately.
2.4.2 Preprocessing step 2: between-gel preprocessing 2.4.2.1 Collation of data vectors into a matrix for each protein fraction Data corresponding to each protein fraction were collated into four matrices, containing data as described in Table 2. www.electrophoresis-journal.com
Electrophoresis 2006, 27, 1288–1299
General
1291
Figure 1. Illustration describing ‘preprocessing step 1’: (a) correction of the 16-bit greyscale images for distortion, and simultaneous within-gel registration of lanes on the gel to selected reference protein bands; (b) identification of lanes and extraction of sample data vectors; (c) normalization to mitigate variation in overall intensity. © 2006 WILEY-VCH Verlag GmbH & Co. KGaA, Weinheim
www.electrophoresis-journal.com
1292
R. Olias et al.
Electrophoresis 2006, 27, 1288–1299
Table 2. Matrix dimensions for each protein fraction Fraction
Number of rows (vectors of lane Matrix didata) in matrix, by sample type mensions Spanish French
Fraction 1 Fraction 2 Fraction 3 Fraction 4
111 112 110 113
103 107 103 103
Total for both countries of origin 214 219 213 216
(2146431) (2196431) (2136431) (2166431)
2.4.2.2 Automated peak registration within each matrix, using a genetic algorithm (GA) and minimum distance approach The second registration step aimed to reduce further variations in band positions unrelated to the sample, which arise from differences between gels in the positions of corresponding bands on the molecular weight axis. A routine was written to carry out registration within each fraction matrix using a GA. Registration was achieved by segmenting each row vector, and shifting, stretching and/ or shrinking the segments (that is, altering the number of elements) using interpolation, with the aim of minimizing the Euclidean distance between the complete transformed row vector and a ‘target’ vector (the mean vector of each fraction). Optimization of the amounts of shifting, stretching and shrinking was achieved by means of a GA. A highly similar approach to registration is described in the recent paper by Forshed et al. [12], where it is used for the alignment of NMR spectra.
2.4.3 Statistical modelling A selection of statistical modelling methods has been used in this work. The ‘data compression’ method of PCA was used to provide an efficient visual description of the complete, partially preprocessed data set. The results of this PCA showed that the greatest source of variance in the data set was protein fraction, and consequently, data from each fraction were treated separately, not only in the second preprocessing step but also in all subsequent analyses.
effects such as extraction repeatability. PLS can be viewed as a data compression method similar to PCA, but in which the scores produced have successively maximized covariance with a criterion variate(s), in this case a binary dummy variate representing the country of origin. Distance-based linear discriminant analysis (LDA) using the Mahalanobis metric was used to obtain a classification model. LDA was applied first to subsets of PLS scores, and subsequently, to subsets of the original variates. Variate selection (or ‘feature selection’) was carried out using a cross-validated GA, as described in Tapp et al. [13]. Variate selection has long been used as a precursor to multivariate analysis, in particular before the advent of the data compression methods. However, a common difficulty has been that in problems where the total number of variates d to choose from is very large, the number of possible subsets is astronomical. (For example, d = 431 in the present work, so the possible number of subsets of, say, size r = 5, is dCr = d!/(r!(d 2 r)!),1011). Nevertheless, attention has turned back to variate selection approaches, since a number of recent studies have shown that they offer better performance than data compression methods [14, 15]. GAs have been found to be effective at selecting variates from high-dimensional data sets [16–18]. In the present work, we have written a GA to search for a small subset of variates to pass to LDA. The criterion for termination of GA was no further improvement in success rate as measured by cross-validation for ten generations of evolution. In view of the random nature of the initial ‘chromosomes’ (variate identifiers), coupled with the possibility of convergence on local optima, the complete GA routine was carried out repeatedly, and the best solution from each execution retained. The frequency of occurrence of each variate calculated across all retained solutions was interpreted as an indication of its usefulness as a discriminator. Note that throughout this work, we have used ‘leave-onesample-out’ cross-validation (that is, the cross-validation segments comprise all data vectors originating from each individual sample in turn). This approach to cross-validation is necessary when there are replicates in the data matrix, to avoid over-fitting which could potentially arise in standard leave-one-out cross-validation.
3 Results and discussion Since an aim of the work was to assess the suitability of protein-gel electrophoresis for food forensic purposes, supervised modelling methods were employed to analyse the data from each fraction, seeking to discriminate systematically between samples from the two different provenances. Partial least squares (PLS) data compression was used for exploring the data sets and examining © 2006 WILEY-VCH Verlag GmbH & Co. KGaA, Weinheim
The effects of the various preprocessing steps are illustrated in Fig. 2 for a single representative sample. Lanes of data taken directly from the corrected gel images (that is, after preprocessing step 1 (after Section 2.4.1.1)) are shown in Figs. 2a and b, respectively, as greyscale images, and as vectors of average greyscale intensity for each www.electrophoresis-journal.com
Electrophoresis 2006, 27, 1288–1299
General
1293
Figure 2. Lanes of data taken directly from the corrected gel images are shown as: (a) greyscale images and (b) as vectors of average greyscale intensity vs. vector element index. Effect of the normalization procedure is illustrated in (c), where each data vector is shown as a ‘reconstructed lane’, and in (d), where the data are plotted as normalized greyscale intensity vs. vector element index. Data are shown after processing with the second registration procedure in (e) and (f).
lane. Figures 2c and d show the effect of the normalization procedure (preprocessing step 1 (after Section 2.4.1.3)). Figure 2d shows the data plotted as normalized greyscale intensity versus vector element index, whereas Fig. 2c shows each data vector as a ‘reconstructed lane’, where each element in the normalized data vector is used as the greyscale intensity of a horizontal line of fixed (but arbitrary) width. This way of presenting the data is used solely to facilitate examination and comparison of the data sets. We see that the normalization step eliminates the issue of gross intensity differences between lanes originating from different gels. The first two scores from the application of PCA to the entire partially pre-preprocessed data set (after preprocessing step 1) are plotted against one another in Fig. 3. It is immediately clear that the fraction is a major source of variance in the data. In the light of this, along with the clear difference between fractions noted on © 2006 WILEY-VCH Verlag GmbH & Co. KGaA, Weinheim
visual examination of the raw data, it was decided to treat the data from each fraction separately in subsequent data analysis. Returning again to Figs. 2c and d, it is evident that an issue of poor registration remains, despite the alignment with the four selected reference protein bands (in preprocessing step 1 (Section 2.4.1.1)). Figures 2e and f show again the data from the same single sample, after the second registration procedure (preprocessing step 2 (Section 2.4.1.2)), which was applied to each fraction separately. The improvement in the alignment of corresponding bands is clearly seen. These figures also give a visual impression of the repeatability of the entire measurement process (extraction process, gel acquisition and imaging). The data from each fraction were explored using PLS. Plots of the first few PLS scores in each case were examined. The plots of the first versus second scores are www.electrophoresis-journal.com
1294
R. Olias et al.
Electrophoresis 2006, 27, 1288–1299
Figure 3. Plot of the first vs. second PC scores obtained from the whole partially preprocessed data set.
shown for all four fractions in Fig. 4. We find that in all cases there is at least some discrimination between the two provenances; fraction 2 shows the greatest distinction between the groups. The first versus second PLS scores for this fraction are shown again in Fig. 5, with the data from a representative selection of samples marked using different symbols; convex hulls for these samples are also drawn on the plot as a visual aid. This plot allows the variance arising from the extraction/measurement procedure to be qualitatively examined (i.e. the within-sample variance) and compared with the variance between samples, and within and between groups of samples. Cross-validated PLS-LDA was carried out, using successively increasing subsets of PLS scores, to determine whether the groups of data could be systematically distinguished. The classification success rate (summarized over the cross-validation segments) is shown versus the number of PLS scores used in Fig. 6. The maximum success rate ranges from ,80% for fraction 1 to ,91% for fraction 2, and in all cases this is achieved using relatively few scores (between 4 and 7). This is an encouraging performance from a laboratory technique that is not generally used in such a quantitative manner. © 2006 WILEY-VCH Verlag GmbH & Co. KGaA, Weinheim
The GA was implemented to identify subsets of variates from each fraction’s data that produce maximum crossvalidation success rate in LDA. Each execution of the GA sought a subset of the same size as the number of scores needed for an optimal model in the PLS-LDA procedure, with the rationale that this represents a minimum number of underlying factors needed to achieve the best possible discrimination. Figure 7 shows histograms depicting the frequency of occurrence of each variate across the 500 GA solutions obtained for each fraction. Inspection of the individual solutions showed that the vast majority comprised representative variates from across several of these bands (rather than several from within one band). The largest features in the histograms are peak-picked and labelled with the corresponding variate identifier; these represent successively the most useful variates. Next, subsets of the peak-picked variates were passed to the cross-validated LDA routine, ranked in order of their occurrence frequency, largest to smallest. The success rate as a function of the number of variates used is shown in Fig. 8. We see that the optimum success rates obtained (ranging from 81 to 91%) are broadly similar to the PLSLDA models, and moreover, from a small number of original variates. This is a very substantial reduction in comwww.electrophoresis-journal.com
Electrophoresis 2006, 27, 1288–1299
General
1295
Figure 4. Plots of first vs. second PLS scores for (a) fraction 1, (b) fraction 2, (c) fraction 3 and (d) fraction 4.
plexity compared with using the entire range of variates as in the PLS-LDA approach. In principle, parsimonious models are desirable because they are more likely to be interpretable in biochemical terms.
Figure 5. First vs. second PLS scores for fraction 2. Selected individual samples are marked with convex hulls and symbols as indicated, to illustrate the experimental reproducibility.
© 2006 WILEY-VCH Verlag GmbH & Co. KGaA, Weinheim
Figure 9 highlights some of the highest-ranked variates (as identified by the histograms) on plots of the means of each group from each fraction. This provides a direct indication of which protein bands can be used to discriminate between samples from the two provenances. It is interesting to note that many of the bands identified are common to more than one fraction; note in particular the correspondence between the variates identified for fractions 1 and 3. The near identity of the variates for these fractions suggests that the same proteins must be involved. It also provides confirmation that the data analysis protocol, which was applied separately to each fraction’s data, is identifying variates that are truly important in distinguishing samples’ provenance and that the solutions identified are neither overfit nor artefacts of, say, the registration procedure. It is interesting to note, however, that the original profiles for fractions 1 and 3 are very different (see Fig. 9), whereas fractions 2 and 3, for instance, resemble each other more (they were extracted using the same buffer). This is also apparent in Fig. 3, where the PC
www.electrophoresis-journal.com
1296
R. Olias et al.
Electrophoresis 2006, 27, 1288–1299
Figure 6. Classification success rate in cross-validated PLS-LDA vs. number of PLS scores used, shown for (a) fraction 1, (b) fraction 2, (c) fraction 3 and (d) fraction 4.
Figure 7. Histograms depicting the frequency of occurrence of each variate across all 500 solutions for each fraction, shown, respectively, for (a) fraction 1, (b) fraction 2, (c) fraction 3 and (d) fraction 4. © 2006 WILEY-VCH Verlag GmbH & Co. KGaA, Weinheim
www.electrophoresis-journal.com
General
Electrophoresis 2006, 27, 1288–1299
1297
Figure 8. Classification success rate in cross-validated GA-LDA vs. number of variates used, shown for (a) fraction 1, (b) fraction 2, (c) fraction 3 and (d) fraction 4.
scores of samples from fractions 1 and 3 show almost no overlap, but the scores for fractions 2 and 3 overlap considerably. We also note that the most important variates identified in fraction 2 do not correspond with those from fraction 3. An explanation is that the GA is capable of selecting minor bands which are nevertheless good discriminators, rather than only the major bands most obvious to the eye. As an addendum to the work presented here, we will seek to identify a selection of the protein bands found to be important discriminators using proteomics techniques. However, the fact that these can be minor features indicates that the removal of selected bands for further analysis will not be completely straightforward in all cases.
4 Concluding remarks At present, we have only attempted to apply the supervised discrimination procedure to distinguish hams from the two countries. The repeatability achieved for individual sample replicates (Fig. 5) suggests that it might be © 2006 WILEY-VCH Verlag GmbH & Co. KGaA, Weinheim
difficult to take the analysis further (e.g. to discriminate between PDO and non-PDO hams) with this particular data set. Nevertheless, in considering the repeatability, the experimental complexity of the combined sample preparation, fractionation and measurement procedure should be recognized. By applying advanced preprocessing and data registration techniques it has proved possible to carry out a quantitative analysis of the data with a rigorous validation procedure. Further, we have identified a small number of band locations that appear to be diagnostic of country of origin. Such a conclusion would have been quite impractical if the traditional qualitative approach to comparing gels had been adopted. This study has been carried out with financial support from the Commission of the European Communities, RTD programme ‘Quality of Life and Management of Living Resources’, project QLK1-2002-02225 ‘Typical Food Products in Europe: Consumer Preference and Objective Assessment (TYPIC)’. It does not reflect the Commission’s views nor anticipate its future policy in this area. www.electrophoresis-journal.com
1298
R. Olias et al.
Electrophoresis 2006, 27, 1288–1299
Figure 9. Most important discriminating variates as identified in the GA-LDA procedure, marked on plots of the means of each group from each fraction. Panels (a)–(d) correspond to fractions 1–4.
5 References [1] Toldrá, F., Flores, M., Crit. Rev. Food Sci. 1998, 38, 331–352. [2] Toldrá, F., Miralles, M.-C., Flores, J., Food Chem. 1992, 44, 391–394. [3] Toldrá, F., Rico, E., Flores, J., J. Sci. Food Agric. 1993, 62, 157–161.
© 2006 WILEY-VCH Verlag GmbH & Co. KGaA, Weinheim
[4] Monin, G., Marinova, P., Talmant, A., Martin, J. F. et al., Meat Sci. 1997, 47, 29–47. [5] Søndergaard, I., Jensen, K., Krath, B. N., Electrophoresis 1994, 15, 584–588. [6] Jensen, K., Søndergaard, I., Skovgaard, I. M., Nielsen, H. B., Electrophoresis 1995, 16, 921–926. [7] Jensen, K., Tygesen, T. K., Kes¸mir, C., Skovgaard, I. M., Søndergaard, I., J. Agric. Food Chem. 1997, 45, 158–161. www.electrophoresis-journal.com
Electrophoresis 2006, 27, 1288–1299
[8] Skarpeid, H.-J., Kvaal, K., Hildrum, K. I., Electrophoresis 1998, 19, 3103–3109. [9] Skarpeid, H.-J., Moe, R. E., Indahl, U. G., Meat Sci. 2001, 57, 227–234. [10] Sánchez-Peña, C., Luna, G., García-González, D. L., Aparicio, R., Meat Sci. 2005, 69, 635–645. [11] Bradford, M. M., Anal. Biochem. 1976, 72, 248–254. [12] Forshed, J., Schuppe-Koistinen, I., Jacobsson, S. P., Anal. Chim. Acta 2003, 487, 189–199.
© 2006 WILEY-VCH Verlag GmbH & Co. KGaA, Weinheim
General
1299
[13] Tapp, H. S., Defernez, M., Kemsley, E. K., J. Agric. Food Chem. 2003, 51, 6110–6115. [14] Walmsley, A. D., Anal. Chim. Acta 1997, 354, 225–232. [15] Shaw, A. D., diCamillo, A., Vlahov, G., Jones, A. et al., Anal. Chim. Acta 1997, 348, 357–374. [16] Leardi, R., Seasholtz, M. B., Pell, R. J., Anal. Chim. Acta 2002, 461, 189–200. [17] Costa, P. A. D., Poppi, R. J., Quimica Nova 2002, 25, 46–52. [18] Goicoechea, H. C., Olivieri, A. C., J. Chem. Inf. Comp. Sci. 2002, 42, 1146–1153.
www.electrophoresis-journal.com
