Professional Documents
Culture Documents
Rapid Prediction of Composition and Flavor Quality of Cheddar Cheese Using ATRFTIR Spectroscopy
A. SUBRAMANIAN, W.J. HARPER, AND L.E. RODRIGUEZ-SAONA
ABSTRACT: Multiple methods are required for analysis of cheese flavor quality and composition. Chromatography and sensory analyses are accurate but laborious, expensive, and time consuming. A rapid and simple instrumental method based on Fourier transform infrared (FTIR) spectroscopy was developed for simultaneous analysis of Cheddar cheese composition and flavor quality. Twelve different Cheddar cheese samples ripened for 67 d were obtained from a commercial cheese manufacturer along with their moisture, pH, salt, fat content, and sensory flavor quality data. Water-soluble components were extracted from the cheese, dried on zinc selenide FTIR crystal and scanned (4000 to 700 cm1 ). Infrared spectra of the samples were correlated with their composition and flavor quality data to develop multivariate statistical regression and classification models. The models were validated using an independent set of ten 67-d-old test samples. The infrared spectra of the samples were well defined, highly consistent within each sample and distinct from other samples. The regression models showed excellent fit (r > 0.92) and could accurately determine moisture, pH, salt, and fat contents as well as the flavor quality rating in less than 20 min. Furthermore, cheeses could also be classified based on their flavor quality (slight acid, whey taint, good cheddar, and so on). The discrimination of the samples was due to organic acids, amino acids, and short chain fatty acids (1800 to 900 cm1 ), which are known to contribute significantly to cheese flavor. The results show that this technique can be a rapid, inexpensive, and simple tool for predicting composition and flavor quality of cheese. Keywords: Cheddar cheese, infrared spectroscopy, rapid analysis
C: Food Chemistry
heese is a major fermented dairy product with a per capita consumption of around 32 pounds per year. It supplies various nutrients such as proteins and vitamins and minerals such as calcium and phosphorus. With an annual production of 3.1 billion pounds, Cheddar cheese is one of the major types of cheeses in the United States (NASS 2007). The composition and flavor of Cheddar cheese are primary determinants of its price and application. Analytical determination of pH, moisture, fat, salt, and flavor serves to evaluate the quality of cheese. Several factors, including raw materials, manufacturing process, nonstarter lactic acid bacteria, cheese type, and the biochemical reactions such as proteolysis, lipolysis, and so on, influence the composition and flavor of cheese (Chen and others 1998; Bachmann and others 1999; Akalin and others 2002; Singh and others 2003). Heterogeneity of cheese, complexity of the maturation process, and interference from matrix compounds complicate the analysis. In addition, simultaneous measurement of flavor-related volatile compounds such as short chain fatty acids and nonvolatile compounds such as amino acids and organic acids is essential. The current methods for cheese analysis, including chromatography and sensory analysis are laborious, time consuming, expensive, and complete characterization may require multiple accessories and methods (Subramanian and others 2009). Also, chromatographic methods generally require separation and concentration that can
be selective for certain classes of compounds. Hence, there is a need for rapid and reliable instrumental method for simultaneous determination of composition and flavor quality of cheese. A rapid method apart from saving time and money for the cheese industry will also help in ensuring better product quality. Various classes of compounds, more importantly water-soluble nonvolatile compounds, and their ratios determine the flavor (Kosikowski and Mocquot 1958; Aston and Creamer 1986; McSweeney and Sousa 2000). Simultaneous monitoring of flavor compounds and the composition requires methods that are capable of monitoring multiple functional groups. Fourier transform infrared (FTIR) spectroscopy is a simple and rapid technique that monitors the molecular vibrations exhibited by various compounds under infrared light (Subramanian and Rodriguez-Saona 2008a). FTIR provides an overall chemical profile of the sample that varies with the composition and the concentration of the compounds. Advances in FTIR instrumentation, development of sampling techniques like attenuated total reflectance (ATR) and chemometrics have made possible to extract information related to composition and conformation of food components from the infrared spectra. Several developments in the field of statistical multivariate analysis have simplified routine analysis of spectroscopic data. Classification tools such as soft independent modeling of class analogy (SIMCA), can be used for characterization of biochemical composition of sample as well as compounds formed during ripening of the cheese. The concentrations of the compounds can be estimated usMS 20080866 Submitted 11/2/2008, Accepted 1/26/2009. Authors are with ing factor analysis methods such as partial least squares regression Dept. of Food Science and Technology, The Ohio State Univ., 2015 Fyffe (PLSR). Court, Columbus, OH 43210, U.S.A. Direct inquiries to author RodriguezApplication of FTIR spectroscopy and multivariate analysis to Saona (E-mail: rodriguez-saona.1@osu.edu). cheese analysis has been presented by several researchers. Initial
Introduction
C292
Absorbance
4000
3600
3200
2800
2400
2000
1600
1200
800
Wavenumber (cm-1)
Figure 1 --- Typical raw (- - -, top) and 2nd derivative ( --- , bottom) FTIR spectra of Cheddar cheese extract. Exactly 7.5 L of the extract were dried on zinc selenide crystal and scanned in the mid-infrared region (4000 to 700 cm1 ). Important functional groups and their region of absorbance are highlighted. Vol. 74, Nr. 3, 2009JOURNAL OF FOOD SCIENCE C293
C: Food Chemistry
Multivariate analyses
The classification and regression analyses of the spectral data R were performed using Pirouette (version 3.11, Infometrix Inc., Woodville, Wash., U.S.A.). The spectra of the cheese samples were mean-centered, derivatized (Savitzky-Golay polynomial filter with a 5-point window) and normalized prior to multivariate analysis. The spectra were then matched with the pH, composition (moisture, salt, and fat) and flavor quality rating data to develop prediction models based on PLSR. A nonlinear iterative partial least squares (NIPALS) algorithm was employed. Classification model to
IR Predicted pH
IR Predicted Score
C: Food Chemistry
38.0 37.5 37.0 36.5 36.0 35.5 35.0 34.5 34.0 34.0
pH
C
1.95
E
11 10 9 8 7 6 5 4 4 5 6 7 8 9 10
Figure 2 --- Partial least squares regression (PLSR) models for prediction of (A) pH, (B) fat, (C) salt, (D) moisture, and (E) avor quality score of cheese samples. The spectra of 67-d-old cheese were transformed into their 2nd derivative, mean-centered, and normalized prior to multivariate analyses.
C294
xtraction with organic solvents enabled removal of compounds that interfered with the detection of essential compounds such organic acids and amino acids that contribute to the flavor (Subramanian and others 2009). Drying of the extract on the crystal resulted in the formation of a uniform film of sample. The drying time per sample was 3 min. FTIR spectra of the extracts were collected in the mid-infrared region (4000 to 700 cm1 ), using a 3bounce zinc selenide ATR crystal. In a 3-bounce ATR-FTIR crystal, infrared light bounces on the sample 3 times, increasing the absorbance and hence the signal. The sample preparation method allowed for the collection of high-quality spectra with distinct spectral features that were very consistent within each sample. The FTIR spectra reflect the total chemical composition of the cheese extract, with absorbance bands due to acids, esters, alcohols, and peptides. The band intensities vary with the overall concentration of the chemical functional groups in the sample. The raw spectra were transformed into their 2nd derivatives to remove the baseline shifts, improve band resolution, and reduce noise and variability between replicates (Kansiz and others 1999). A typical FTIR spectrum of Cheddar cheese extract and its 2nd derivative are shown in Figure 1. The region from 4000 to 3100 cm1 consists of absorbance from OH and NH stretching vibrations of hydroxyl groups and Amide A of polypeptides and amino acids, respectively. The CH stretching vibrations of CH 3 and > CH 2 functional groups of fatty acids appear between 3100 and 2800 cm1 . The spectral range 1800 to 900 cm1 contains signal from polypeptides, amino acids, carbonyl groups of fatty acids, hydroxyl groups, carboxylic acid groups, and fatty acid esters (typically short chain). Visual comparison of the raw spectra showed numerous differences between cheeses, especially in the spectral region 1800 to 900 cm1 . The transformed spectra were correlated with the composition (fat, salt, and moisture), pH, and quality rating and analyzed by PLSR with cross-validation (leave-one-out approach) to generate calibration models. PLSR is a bilinear regression analysis which determines the analytes concentration (Y-variable) by regressing a small number of orthogonal factors that are linear combinations of variables (X-variable; infrared wavenumbers). These orthogonal
2 5 4 6
PC3
8 7
PC1 PC2
Figure 3 --- Soft independent modeling of class analogy (SIMCA) classication plot for discrimination of cheese samples based on avor quality: 1 and 2: both slight acid and good creamy, 3 and 4: both slight whey taint and medium creamy, 5 and 6: both good creamy cheddar, and 7 and 8: low avor and good creamy. The samples were projected against the rst 3 principal components (PC) that explained the largest amount of variance among the samples.
Table 1 --- Predicted and actual pH, fat, moisture, salt, and avor quality score for the 67-d-old test samples. Test sample A B C D E F G H I J pH Predicted Actual 5.18 5.13 5.10 5.14 5.11 5.08 5.11 5.13 5.20 5.23 5.18 5.12 5.11 5.17 5.11 5.07 5.10 5.10 5.19 5.22 Fat (%) Predicted Actual 34.79 35.03 35.28 35.20 35.19 35.07 35.07 35.06 35.17 34.88 34.80 34.40 35.40 35.10 35.50 35.20 35.20 35.00 35.40 34.60 Moisture (%) Predicted Actual 35.77 35.65 35.45 35.59 35.71 35.78 35.47 35.86 35.86 36.23 35.80 36.20 35.50 35.60 35.50 35.60 35.40 35.80 35.60 36.50 Salt (%) Predicted Actual 1.83 1.82 1.77 1.79 1.80 1.78 1.76 1.80 1.76 1.81 1.86 1.84 1.78 1.83 1.78 1.73 1.75 1.78 1.74 1.81 Flavor quality score Predicted Actual 6.02 6.17 6.93 7.49 7.36 8.81 7.65 7.75 6.20 4.88 6.00 6.00 7.00 8.00 7.00 9.00 8.00 8.00 6.00 5.00 C295
C: Food Chemistry
2000 Organic Acids (C-O and C-C Stretching Vibrations Esters of Fatty 1000
1500
Figure 4 --- Discriminating power plot for classication of Cheddar cheese samples. The regions of the FTIR spectra that contributed to the discrimination of the cheese samples based on their avor are highlighted. Higher the discriminating power at a particular wavenumber the greater is the difference between the samples in the chemical groups associated with that wavenumber.
C: Food Chemistry
500
0 850
1050
1250
1450
1650
1850
Wavenumbers (cm-1)
The PLSR prediction models described previously were validated with an independent set of ten 67-d-old test samples. The pH, fat, salt, moisture content, and the flavor quality score were predicted using the developed models. The values predicted using the models and the actual values are presented in Table 1. The models showed excellent predictive capability. The average percentage deviation of predicted values from actual values were 0.23, 0.57, 0.47, 1.23, and 3.1 for pH, fat, moisture, salt, and flavor quality score, respectively. These data clearly indicate the reliability of the models and the potential of this technique for simultaneous analysis of cheese characteristics. The spectra were also analyzed by the soft independent modeling of class analogy (SIMCA) with the aim of classifying the cheese based on their flavor notes. SIMCA is a multivariate statistical technique based on principal component analysis (PCA) that reduces the dimensionality of multivariate data sets. In SIMCA, training sets are assigned to classes and a principle component model is generated for each class with distinct confidence regions within them (De Maesschalck and others 1999). A score plot is constructed by projecting the actual data on the first 3 principal components that explain the most variance in the training data set. The variance explained by the class model describes the signal and the residual variance describes the noise in the data set and is used to define probability boundaries around the class. Comparison of the average residual variance of samples in a class and the residual variance of an unknown sample can help in identification of the unknown sample (Lavine 2000). The performance of this method depends not only on the difference between classes, but also strongly on the training set for each class (Candofi and others 1999). The SIMCA classification plot for discrimination of cheese based on flavor is shown in Figure 3 (only 8 of the 12 samples are shown for ease of visualization). All 8 samples formed tight clusters and the location of the clusters in 3D space correlated well with their flavor quality. The good samples clustered together (clusters 5 and 6) and away from the samples with defects. Clusters with defects in flavor notes clustered away from the good creamy cheddar but close to samples with similar flavor note. The distance between the clusters in a SIMCA plot is represented by the interclass distance (ICD). It is a measure of the quality of
C296 JOURNAL OF FOOD SCIENCEVol. 74, Nr. 3, 2009
the data. Greater the distance between 2 clusters the greater is the difference in composition of samples belonging to those clusters. As a rule of thumb, a distance of over 3 indicates that the samples are well separated (Kvalheim and Karstand 1992). Samples that had similar flavor quality had ICD of less than 3 within themselves and an ICD of greater than 3 when compared to samples with a different flavor quality (data not shown). These data support our previous publication on the possibility of classifying cheese samples based on flavor notes (Subramanian and others 2009). The spectral wavenumbers and the associated functional groups that were responsible for the classification of the cheeses in SIMCA plot can be identified using the discriminating power plot. In the discriminating power plot each wavenumber in the spectral range is plotted against its importance in discriminating the samples that are in the model. The higher the value of discriminating power, the greater is the influence of that wavenumber in classifying the samples. The discriminating power also indicates the quality of the data and variables with low discriminating power are usually deleted because they contribute only to noise in the data set (Lavine 2000). The spectral regions and the associated functional groups/compounds responsible for the differentiation of the Cheddar cheese samples are highlighted in Figure 4. The spectral range 1800 to 900 cm1 was found to be important in the analysis of cheese flavor by FTIR. This region consists of signals from CO and C=O (approximately 1175 cm1 ), CH bending (approximately 1450 cm1 ), esters (1750 to 1700 cm1 ), and CO stretching (approximately 1240 and 1170 to 1115 cm1 ) (Rodriguez-Saona and others 2006). In the case of Cheddar cheese extract compounds containing these functional groups include the organic acids, alcohols, short chain fatty acids and their esters, amino acids, and small water-soluble peptides, all of which are important for flavor.
rapid, simple, and reliable FTIR technique was developed for simultaneous analysis of Cheddar cheese composition and flavor quality. The fat, salt, moisture, pH, and flavor quality score of the cheese could be predicted in less than 20 min. Additionally, the infrared spectra correlated well with the flavor notes. This technique could be a rapid quality control tool for the cheese industry to
Conclusions
Acknowledgments
This study was supported by the USDA Cooperative State Research, Education and Extension Service, special research grant nr 200634328-17149.
References
Adams MJ, Latham K, Barnett NW, Poynton AJ. 1999. Calibration models for determining moisture and fat content of processed cheese using near-infrared spectrometry. J Sci Food Agric 79:12326. Akalin AS, Gonc S, Akba Y. 2002. Variation in organic acids content during ripening of pickled white cheese. J Dairy Sci 85:16706. Aston J, Creamer LK. 1986. Contribution of the components of water-soluble fraction to the flavor Cheddar cheese. New Zealand Dairy Sci Tech 21:22948. Bachmann HP, Butikofer U, Meyer J. 1999. Prediction of flavor and texture development in Swiss-type cheeses. Lebensm-Wiss U-Technol 32:2849. Bjorsvik HR, Martens H. 1992. Data analysis: calibration of NIR instruments by PLS regression. In: Burns DA, Ciurczak EW, editors. Handbook of near-infrared analysis. New York: Taylor and Francis. p 15980. Blazquez C, Downey G, OCallaghan D, Howard V, Delahunty C, Sheehan E, Everard C, ODonnell CP. 2006. Modelling of sensory and instrumental texture parameters in processed cheese by near infrared reflectance spectroscopy. J Dairy Res 73:5869. Candofi A, De Maesschalck R, Massart DL, Hailey PA, Harrington ACE. 1999. Identification of pharmaceutical excipients using NIR spectroscopy and SIMCA. J Pharma Biomed Anal 19:92335. Cattaneo TMP, Giardina C, Sinelli N, Riva M, Giangiacomo R. 2005. Application of FTNIR and FTIR spectroscopy to study the shelf-life of Crescenza cheese. Int Dairy J 15:693700. Chen M, Irudayaraj J, McMahon DJ. 1998. Examination of full fat and reduced fat Cheddar cheese during ripening by Fourier transform infrared spectroscopy. J Dairy Sci 81:27917. De Maesschalck R, Candofi A, Masart DL, Heuerding S. 1999. Decision criteria for soft independent modeling of class analogy applied to near infrared data. Chemometr Intell Lab Syst 47:6577. Downey G, Sheehan E, Delahunty C, OCallaghan D, Guinee T, Howard V. 2005. Prediction of maturity and sensory attributes of Cheddar cheese using near-infrared spectroscopy. Int Dairy J 15:7019. Kansiz M, Heraud P, Wood B, Burden F, Beardall J, McNaughton D. 1999. Fourier transform infrared microspectroscopy and chemometrics as a tool for the discrimination of cyanobacterial strains. Phytochemistry 52:40717.
C297
C: Food Chemistry