Multivariate Curve Resolution of Spectrophotometric Data for the Determination of Artificial Food Colors DIRK W. LACHENMEIER* ,† AND WALTRAUD KESSLER ‡ Chemisches und Veterina ¨runtersuchungsamt (CVUA) Karlsruhe, Weissenburger Strasse 3, D-76187 Karlsruhe, und Institut fu ¨r Angewandte Forschung, Reutlingen University, Alteburgstrasse 150, D-72762 Reutlingen, Germany In the analysis of food additives, past emphasis was put on the development of chromatographic techniques to separate target components from a complex matrix. Especially in the case of artificial food colors, direct spectrophotometric measurement was seen to lack in specificity due to a high spectral overlap between different components. Multivariate curve resolution (MCR) may be used to overcome this limitation. MCR is able to (i) extract from a complex spectral feature the number of involved components, (ii) attribute the resulting spectra to chemical compounds, and (iii) quantify the individual spectral contributions with or without a priori knowledge. We have evaluated MCR for the routine analysis of yellow and blue food colors in absinthe spirits. Using calibration standards, we were able to show that MCR equally performs as compared to partial least-squares regression but with much improved chemical information contained in the predicted spectra. MCR was then applied to an authentic collective of different absinthes. As confirmed by reference analytics, the food colors were correctly assigned with a sensitivity of 0.93 and a specificity of 0.85. Besides the artificial colors, the algorithm detected a further component in some samples that could be assigned to natural coloring from chlorophyll. KEYWORDS: Multivariate curve resolution; MCR; PLS; spectrophotometry; food colors; alcoholic beverages; absinthe INTRODUCTION Chemometric methods such as principal component analysis (PCA) and partial least-squares (PLS) regression have been used successfully in many applications over the years (1–3). All of these applications use the multivariate methods to reduce the multidi- mensional data sets to fewer dimensions and to point out inter- correlations and interdependencies in the data. The majority of applications in chemistry are within the field of spectroscopy. The disadvantage of a PCA or PLS approach is that the obtained principal components are abstract mathematical factors, so-called “latent variables”, with usually little or no physical or chemical meaning. The regression coefficient of the PLS sometimes provides hints to attribute defined features in the spectra to the response variable. However, many users, especially in an industrial environ- ment, want to obtain information that is as close as possible to their real life experience in spectroscopy. Much research has been done to solve the mixture analy- sis problem and to extract real spectra and concentration profiles from overlapping spectral data without any a priori assumptions about the composition of the system. Several mixture analysis methods are known such as evolving factor analysis (EFA) (4), fixed-size moving window evolving factor analysis (FSMWEFA) (5), target factor analysis (TFA) (6), classical curve resolution (CCR) (7), weighted curve resolution (WCR) (8), multivariate curve resolution (MCR) (9–12), and to a certain extent also techniques such as parallel factor analysis (PARAFAC) (13). In food chemistry and especially the analysis of food additives, emphasis in the past has been on the development of separation methods to analyze the target compounds as selec- tively as possible. Food colors are regularly analyzed using thin- layer chromatography (TLC) or high-performance liquid chro- matography (HPLC) (14–24). The direct spectroscopy of the original food matrix without separation was not possible in the past, because the resulting spectra are difficult to interpret and often lack specificity. This disadvantage can be solved using MCR, as it implies the following objectives: 1. Resolve the number of chemical compounds simulta- neously present in the mixture from a complex spectral signature. 2. Identify these species by transforming mathematical solu- tions to real spectra, thus increasing specificity by applying mathematical and chemical constraints. 3. Quantify each component without any prior assumption or knowledge of the chemical model involved. Unlike deconvolution, MCR provides spectra of pure com- pounds and not only resolution of single bands, which are difficult to attribute to chemical compounds in a complex * To whom correspondence should be addressed. Tel: +49-721-926- 5434. Fax: +49-721-926-5539. E-mail: Lachenmeier@web.de. † CVUA Karlsruhe. ‡ Reutlingen University. J. Agric. Food Chem. 2008, 56, 5463–5468 5463 10.1021/jf800069p CCC: $40.75  2008 American Chemical Society Published on Web 06/24/2008