Food Science and Technology Research
Online ISSN : 1881-3984
Print ISSN : 1344-6606
ISSN-L : 1344-6606
Original papers
Coupled Stepwise PLS-VIP and ANN Modeling for Identifying and Ranking Aroma Components Contributing to the Palatability of Cheddar Cheese
Airi Morita , Tetsuya Araki, Shoma Ikegami, Misako Okaue, Masahiro Sumi, Reiko Ueda, Yasuyuki Sagara
Author information
Keywords: partial least-squares regression-variable importance in projection (PLS-VIP), artificial neural network (ANN), gas chromatography / olfactometry (GC/O), fermented food, food kansei model
JOURNAL FREE ACCESS FULL-TEXT HTML

2015 Volume 21 Issue 2 Pages 175-186

Details
Download PDF (2426K)
Download citation RIS

(compatible with EndNote, Reference Manager, ProCite, RefWorks)

BIB TEX

(compatible with BibDesk, LaTeX)

Text
How to download citation
Contact us
Abstract

A consumer-oriented methodological approach for the quality evaluation of Cheddar cheese as a typical fermented food was developed. Datasets were obtained from gas chromatography/olfactometry (GC/O) analysis and sensory evaluation of 10 Cheddar cheese samples. The GC/O analysis identified 43 aroma components under the categories of 14 aroma descriptors. Consumer evaluation of palatability was performed by 59 housewives. Factor analysis of the GC/O data identified aroma descriptors that have positive or negative correlations with palatability scores. Twelve aroma components were prioritized using stepwise partial least-squares regression with variable importance in projection (PLS-VIP). An artificial neural network (ANN) model was constructed to demonstrate the nonlinear relationships among the raw GC/O data of the samples and the palatability scores. Coupling stepwise PLS-VIP and ANN resulted in successful identification and ranking of aroma components contributing to the palatability of Cheddar cheese, and in modeling their nonlinear relationships.

Introduction

A consumer-oriented kansei engineering approach is increasingly needed to support new product development in the food industry. Kansei engineering was founded 40 years ago and was originally developed as a methodology for transforming a consumer's feeling or image about a product into the design elements of the product (Nagamachi, 1995). Subsequently, Sagara (1994) proposed food kansei engineering, a new area of scientific and engineering research for conducting quantitative measurements of the sense of taste and food preferences by the systematization of individual techniques, such as instrumental analyses, sensory evaluation, and multivariate analyses. The methodological combination of quantitative data and marketing tools aids in the formulation of improved strategies for product development (Akiyama et al., 2012).

By applying the theoretical framework of food kansei engineering, Ikeda et al. (2004) proposed a food kansei model for quantitatively evaluating personal experiences of gustatory sensations during short-term eating behavior, and for utilizing the quantitative information to optimize product design and manufacturing processes. The model has been applied to investigate the relationships between instrumental and sensory data of bottled green tea beverages (Ikeda et al., 2004), sesame-flavored dressings (Ikeda et al., 2006), Danish pastries (Shibata et al., 2008), prepackaged chilled espresso beverages (Michishita et al., 2010), and ready-to-drink coffee beverages (Akiyama et al., 2012). However, applying the model to fermented foods has remained a challenge due to their unique flavor and texture.

In general, due to their unique flavor and texture, fermented foods are often disliked by people of different cultures. Historically, processed cheeses have been more readily accepted by the Japanese compared to natural cheeses; the primary reason being that that several decades ago, the relatively tasteless flavor and texture of processed cheeses made them preferable to natural cheeses. In particular, at the time, the unique flavor and texture of natural cheeses were not well accepted among the Japanese; the annual cheese consumption per capita was less than 10 g until 1950. Decades later, overall changes in Japanese dietary habits, illustrated by the consumption of pizza, cake, and wine, were accompanied by increased consumption of natural cheeses.

In recent years, the total cheese consumption of 38 selected countries reached 1.69 million metric tons in 2012; the consumption for USA and Japan was 6,585 and 301 thousand metric tons, respectively (USDA, 2013). Despite the comparatively smaller cheese consumption of Japan, there is an increasing trend in the consumption of natural cheese products. In 2012, the total cheese consumption in Japan was 301,495 metric tons, a 6% increase over the previous year, while natural cheese consumption was 181,354 metric tons, a 12.8% increase over the previous year (MAFF, 2013). In addition, natural cheeses account of approximately 96.0% of total cheese imports in Japan, and are not only directly consumed but also utilized as ingredients of processed cheeses, mainly made from Cheddar and Gouda cheeses.

Since aroma and flavor are important factors influencing food palatability, the relationships between measured aroma components and sensory evaluation scores were also discussed based on these datasets in several previous studies (Carunchia Whetstine et al., 2005; Carunchia Whetstine et al., 2006) and the results of statistical analysis (Kim et al., 2011; Thomsen et al., 2012). In addition, there has been extensive research on the flavor of Cheddar cheese and other varieties, but despite this effort, only limited information is available on the flavor chemistry of most varieties, and none of the flavors are characterized sufficiently to permit reproduction by mixtures of pure compounds in a model cheese (Fox et al., 1995; Parliament and McGorrin, 2000; McGorrin, 2001; Singh et al., 2003).

Furthermore, many previous studies have been carried out to explore important variables from large amounts of information such as spectral data, mainly in the field of chemometrics. Factor analysis is often used, assuming the presence of potential common factors (Roy and Roy, 2009; Saurina, 2010). On the other hand, since many variables in the datasets of organic composition in food are often collinear, partial least-squares regression (PLS) is applied as a method to overcome the problem of multicollinearity and for data reduction. Then, PLS, a linear regression method, and other nonlinear regression models are usually combined to obtain the final discriminating results (Bhandare et al., 1993; Wang et al., 1999; Wang et al., 2003; Ciosek et al., 2005). With the growing appreciation of nonlinear phenomena in general, the artificial neural network (ANN) is the most promising candidate for tractable interpretation and modeling of a large class of nonlinear problems (Borggaard and Thodberg, 1992). The ANN was employed to investigate the relationships between the principal component scores of data obtained from GC/O analysis and comprehensive palatability scores of beverage samples such as bottled green tea (Ikeda et al., 2004) and prepackaged chilled espresso (Michishita et al., 2010). However, in these studies, raw GC/O data of the aroma components have not been directly applied to ANN modeling. At any rate, research methods in the above-mentioned previous studies were unable to fully investigate all the relationships among the datasets obtained from instrumental analyses and consumer palatability scores, making it difficult to identify important aroma components relevant to the comprehensive palatability of natural cheeses. In this context, methodological developments are required to identify the aroma components contributing to consumer acceptance, especially the comprehensive palatability scores evaluated by consumers.

The objectives of this study were to develop a consumer-oriented methodological approach for the quality evaluation of Cheddar cheese as a typical fermented food, based on the food kansei model and the relationships between the data obtained from instrumental analyses of aroma and consumer palatability evaluations, and to identify characteristic aroma components of Cheddar cheese and the degree of these contributions to the comprehensive palatability scores evaluated by Japanese consumer panelists using coupled partial least-square regression - variable importance in projection (PLS-VIP) and ANN modeling.

Materials and Methods

Samples and aroma extraction    Ten commercial Cheddar cheeses labeled as samples No.1 to 10 were selected as test samples.

Samples were commercially available (No.3, 9, and 10) and ingredients of processed cheeses (No.1, 2, 4, 5, 6, 7, and 8) (Table 1). The basic model was formulated using samples with a normal aroma and not samples with anomalous characteristics. Thus, the samples were selected from normal Cheddar cheese, indicating that samples have normal characteristics in their aroma or flavor and are normally useable for both direct consumption and as an ingredient for processed cheeses.

Table 1. Test sample data
No. Area of Production Date of Manufacture Aging period**
1 Oceania 2010.11.24 5
2 USA 2011.03.17 2
3 Japan 2011.07.20 4
4 Oceania 2010.12.10 6
5 Oceania 2011.02.07 4
6 Oceania 2012.03.05 0
7 USA 2011.11.17 -
8 Oceania 2011.01.-* 5
9 Japan - -
10 Europe 2012.03.27 5
*  Date was unknown,

**  Unit; month

Aromas of the samples were analyzed using GC/MS and GC/O to identify their characteristic components and intensities. Aroma components were extracted from the samples (20 mm cubic columns) by using solvent assisted flavor evaporation (SAFE). Each sample was frozen in liquid nitrogen and then ground with an Osterizer blender (Jarden Consumer Solutions, Boca Roton, FL, USA). A 300-mL aliquot of a dichloromethane solution of 3-heptanol (20 µg/mL) and 4-octanol (1.0 µg/mL) was added to the powdered sample as a standard substance, and the volatiles were extracted for 8 h using a Soxhlet apparatus (Engel et al., 1999). The extract was concentrated to about 200 mL before being distilled with the SAFE apparatus. The obtained distillate was concentrated in vacuo by evaporation to about 1.0 mL (Onishi et al., 2011).

GC/MS and GC/O analyses    The concentrated extracts were analyzed with a 7890A gas chromatograph equipped with a 5975B mass spectrometer (Agilent Technologies Inc., Palo Alto, CA, USA). A DB-WAX capillary column (60 m × 0.25 mm, 0.25 µm film thickness, Agilent Technologies) was employed and the helium carrier gas flow rate was maintained at 1.6 mL/min. The oven temperature was programmed at an initial 50°C for 2 min, and was increased at 3°C/min to 220°C, and then held at the constant temperature for 75 min. The injection port was maintained at 250°C (25 psi). The extract (1.0 µL) was injected into pulsed split mode (50 psi; 60 s) to minimize thermal degradation in the injection port (Michishita et al., 2010).

GC/O analysis of the concentrated extract and the dichloromethane solutions with dilution ratios of 1/3 and 1/9 were conducted using Charm AnalysisTM with an Agilent 6890GC (modified by DATU, Inc., Geneva, NY, USA) (Acree et al., 1984). A DB-WAX fused-silica capillary column (15 m × 0.32 mm, 0.25 µm film thickness; Agilent Technologies) was employed with a helium carrier gas flow rate of 3.2 mL/min. The oven temperature was programmed at an initial 40°C, increased at 6°C/min to 230°C, and then held at 230°C for 20 min (Michishita et al., 2010). The injection port and detector were maintained at 225°C, and a 1-µL sample was injected in split-less mode. The GC effluent gas of each sample was sniffed in humidified air by a trained panelist. The odor activities of volatile components obtained by GC/O dilution analyses were represented as charm value (CV) as well as aroma descriptors (Acree et al., 1984). The CV was an index of odor intensity and was calculated by integrating the time length and the dilution level as follows:   

where F: dilution factor, n: number of dilutions, and di: length of time.

The volatile components were identified by comparing their mass spectra and Kovats indices with those of standard components. Some potent odorants found only by the GC/O analysis were tentatively identified by comparing their Kovats indices and aroma properties with those of standard components (Ikeda et al., 2006; Onishi et al., 2011).

Consumer evaluation of palatability    Consumer evaluation of palatability was carried out by 59 consumers (housewives) in their 30s, 40s, or 50s. All panelists were housewives residing in the neighborhood where the consumer evaluation was conducted. The panelists did not receive any form of training regarding dairy products or cheeses. Female housewives were chosen as the target consumer because they were considered to have the greatest influence on daily purchases among family members in Japan regarding cheese products. Consumer panelists were selected using preliminary questionnaires on food allergy and personal preferences in regard to dairy products and cheese. The panelists gathered prior to lunch (between 10:30 am and 11:00 am), and their physical conditions and food allergies were checked before consumer evaluation. As a result, appropriate panelists were employed for consumer evaluation.

The samples were cut into 15-mm cubes, placed in plastic cups labeled with a three-digit code, and then provided to panelists. The consumer palatability tests were carried out at room temperature with fluorescent lamp lighting. The panelists tested each sample at white desks arranged in the same testing room. The panelists were allowed to carry out both finger and sniffing tests, followed by swallowing the sample. A nine-point category scale was used to evaluate consumer palatability with a hedonic rating system.

Statistical analyses    Obtained datasets, including the aroma components and consumer palatability scores, were subjected to analysis of variance (ANOVA), factor analysis, correlation analysis, partial least-squares regression (PLS) and artificial neural network (ANN) modeling using JMP 9.0 (SAS Institute, Inc., Cary, NC, USA).

Factor analysis was carried out for total aroma intensity values of 14 aroma descriptors obtained by GC/O using the maximum-likelihood method and Varimax rotation for the factors in which the variance was over 1.0.

In practice, especially in process engineering, a method is required to identify the explanatory variables with the most influential on the quality of the final product. However, as mentioned earlier, many variables in the datasets of food products are often collinear. Currently, the variable importance in projection (VIP) scores obtained by PLS regression are paid increasing attention as an important measure of each explanatory variable or predictor (Chi-Hyuck et al., 2009). The stepwise application of PLS-VIP was utilized in identifying the components conforming consumer palatability. Although the coefficients of determination (R2) of PLS would decrease during the components selection process, the order determination of specified components appeared to be more important in practical aspects. Therefore, the VIP and the weight of model (W) were monitored as threshold indicators to determine the contribution degree of each component. Thresholds for cutoff value in PLS-VIP were set as VIP > 0.80 and approximately W = ± 0.1 according to the default setting of the software (Wold et al., 1984).

ANN modeling was carried out to evaluate the link among selected aroma components and comprehensive palatability. The network had the architecture of a characteristic tri-layer perception network. This consists of one hidden layer between the input and output layers, modules of which use an error back propagation algorithm for weight to prediction error. The number of nodes in the input layer of each network was equivalent to the specified aroma components obtained from PLS-VIP. The number of hidden layers was determined by the minimum number required to improve study accuracy using four nodes. In the output layer, the number of nodes was the same as the number of equivalent palatability. The ANN model was trained with maximum iterations of 100 and number of tours of 50 times. The over-fit penalty condition was set to 0.001, and the convergence criterion was set to 0.00001. The coefficients of determination (R2) and root mean square error (RMSE) were used to evaluate the prediction accuracy of the model.

Results and Discussion

Aroma components of Cheddar cheese    GC/MS analysis identified 98 aroma components using their retention index (Ri) and density (mg/100 g) (Table 2). GC/O detected 43 aroma components (Table 3), among which, 15 aroma components corresponded to the Cheddar cheese aroma found in the literature (Singh et al., 2003). Nineteen aroma components were detected by GC/O alone and not by GC/MS. The following 12 aroma components were simultaneously detected by GC/MS and GC/O: acetic acid, butyric acid, decanoic acid, diacetyl, hexanoic acid, p-cresol, phenylacetaldehyde, γ-dodecalactone, δ-decalactone, (Z)-6-dodecen-4-olide, methional and 2,5-dimethyl-4-hydroxy- 3(2H) furanone. The aroma composition of sample No.4 was governed by ‘potato-like’ aroma (methional; 57.2% of total CVs). Similarly, sample No.5 was governed by ‘sulfur’ (dimethyl disulfide; 24.4% of total CVs), and No.6 by ‘floral/fruity’ (ethyl cinnamate; 38.2% of total CVs). From these results, samples No.4, No.5 and No.6 were discriminated from other samples by their strong character, dominated by a single aroma component (See Table 3).

Table 2. Aroma components identified using GC/MS
Peak no. Components/ Sample No. 1 No. 2 No. 3 No. 4 No. 5 No. 6 No. 7 No. 8 No. 9 No. 10
1 diacetyl 5.27* 1.09 3.13 0.86 1.43 0.57 10.78 0.84 0.36 0.90
2 2-pentanone 2.09 0.82 3.72 3.80 4.53 2.66 3.22 0.58 2.08 1.29
3 valeraldehyde 1.86 0.58 1.35 1.44 2.58 0.78 0.84 1.48 0.58 5.60
4 methyl butyrate 0.49 0.30 0.85 0.52 0.62 0.60 0.64 0.24 1.27 0.95
5 3-methyl-3-buten-2-one 1.45 1.11 1.02 1.42 1.30 0.29 0.36 0.25 0.13 0.30
6 α-pinene 0.58 0.11 0.17 0.83 1.46 1.60 0.28 0.80 0.06 0.18
7 ethyl butyrate 1.67 8.87 2.34 4.39 4.09 1.82 0.89 4.20 4.34 1.76
8 hexanal 0.70 1.25 0.84 0.61 0.55 2.59 1.91 1.09 1.43 1.28
9 isobutanol 0.19 N.D.** N.D.** 0.19 N.D.** N.D.** N.D.** N.D.** N.D.** N.D.**
10 beta-pinene 0.66 0.36 0.37 1.30 1.95 0.45 1.09 0.59 0.10 0.92
11 2-pentanol 0.12 0.06 0.14 N.D.** 0.03 N.D.** N.D.** 0.43 N.D.** N.D.**
12 ethyl valerate N.D.** 0.15 N.D.** N.D.** N.D.** N.D.** N.D.** N.D.** N.D.** N.D.**
13 butanol 0.17 0.30 N.D.** 0.11 2.85 N.D.** N.D.** N.D.** N.D.** N.D.**
14 myrcene 0.06 3.41 0.74 0.15 0.36 0.09 1.21 0.35 0.02 0.92
15 3-penten-2-ol 3.41 2.40 5.25 3.51 2.87 2.99 1.87 3.68 2.64 1.59
16 2-heptanone 5.87 2.74 13.90 7.65 16.10 5.44 6.20 3.21 6.62 4.30
17 methyl hexanoate 0.71 0.34 1.20 0.66 0.69 0.92 0.34 0.71 0.31 0.87
18 3-methyl-2-butenal N.D.** 0.03 N.D.** 0.33 0.18 N.D.** N.D.** N.D.** N.D.** N.D.**
19 limonene 3.36 4.67 5.47 3.68 4.04 8.12 48.92 17.41 1.01 43.49
20 isoamyl alcohol N.D.** 0.21 0.48 0.34 1.21 0.08 1.13 N.D.** 0.41 9.58
21 ethyl hexanoate 2.50 8.02 1.84 2.98 2.96 0.55 0.99 4.85 1.68 0.35
22 3-methyl-3-butenol 1.95 0.49 1.22 0.83 0.47 0.52 1.29 0.84 1.71 0.94
23 amyl alcohol 1.43 0.43 0.51 1.56 1.57 1.01 0.89 1.71 0.25 4.58
24 p-cymene 0.14 0.11 0.09 0.15 0.43 5.51 0.90 2.00 0.08 0.79
25 acetoin 1233.00 480.02 1088.84 104.46 279.43 166.68 161.98 148.07 56.54 215.22
26 acetol 39.27 7.58 15.58 14.07 20.00 8.47 44.93 11.71 12.28 1 0.3 1
27 methyl lactate 1.26 1.58 2.76 1.34 2.21 0.94 2.13 1.00 1.54 1.87
28 2,6-dimethylpyrazine 3.28 3.20 0.33 0.53 0.14 2.32 1.13 4.61 0.79 1.40
29 ethyl lactate 0.49 6.51 0.83 0.65 1.80 0.11 2.57 1.19 2.70 0.50
30 3-hydroxy-3-methyl-2-butanone 0.57 9.41 16.54 0.20 0.87 0.38 5.87 0.37 6.30 8.51
31 hexanol 0.50 1.24 1.53 0.25 5.61 0.27 0.60 0.48 0.65 0.44
32 2-hydroxy-3-pentanone 0.65 3.96 4.85 0.29 0.85 0.25 6.11 0.51 0.75 7.17
33 2-nonanone 1.80 1.71 7.75 1.98 5.06 1.82 2.06 2.41 2.54 1.50
34 nonanal 0.30 0.42 0.28 0.15 0.13 0.31 0.38 0.54 0.43 0.57
35 trimethylpyrazine 0.49 0.77 1.03 0.05 0.04 0.17 0.18 0.33 0.16 0.21
36 ethyl octanoate 0.37 3.67 0.97 0.49 0.43 0.16 0.25 2.27 0.64 0.17
37 acetic acid 34.48 19.57 417.22 45.76 139.77 25.77 98.71 29.65 74.53 115.39
38 methional 0.32 0.82 N.D.** 1.67 0.44 N.D.** 0.09 N.D.** 0.03 0.19
39 α-copaene N.D.** N.D.** N.D.** 0.01 0.05 N.D.** N.D.** N.D.** N.D.** N.D.**
40 benzaldehyde 0.43 0.35 1.25 0.77 0.17 0.30 0.62 1.17 0.24 1.22
41 tetrahydro-2-methylthiophen-3-one N.D.** 0.31 N.D. N.D. N.D. N.D. 0.13 N.D. 0.43 0.29
42 2-(methylthio)ethanol 0.21 0.07 N.D. 0.20 0.17 N.D. 0.36 N.D. 1.65 0.16
43 propionic acid 2.03 1.27 3.15 1.29 1.73 0.89 1.81 0.96 1.05 1.84
44 2,3-butanediol 8.29 2.39 1736.53 2.90 20.65 2.40 2.65 33.75 174.35 21.88
45 octanol 0.26 0.23 4.55 0.20 0.49 0.17 0.21 0.28 1.71 0.48
46 2,3-butanediol 3.31 1.89 536.14 266.89 165.03 204.47 251.25 366.88 240.47 232.41
47 dimethyl sulfoxide 0.54 3.17 N.D. 0.75 N.D. 0.12 0.05 0.08 0.29 0.02
48 2-undecanone 0.61 0.55 2.65 0.47 0.79 0.79 0.65 0.97 1.12 0.49
49 β-caryophyllene 0.05 0.11 N.D. 0.10 0.42 N.D. N.D. N.D. N.D. N.D.
50 γ-valerolactone 0.32 0.19 0.57 0.17 0.14 0.21 0.19 0.22 0.17 0.17
51 butyric acid 408.65 419.05 1338.98 584.13 703.92 394.45 412.24 449.27 535.03 616.33
52 ethyl decanoate 0.58 4.33 1.50 0.68 0.34 0.14 0.21 2.29 0.77 0.11
53 phenylacetaldehyde 1.63 2.79 6.24 2.88 0.55 1.53 3.37 4.32 0.81 3.04
54 furfuryl alcohol 0.25 0.06 0.44 0.62 0.43 0.04 0.25 0.12 0.14 0.44
55 nonanol 0.18 0.08 0.33 0.06 N.D. 0.12 0.13 0.28 0.15 0.19
56 isovaleric acid 0.68 0.74 1.69 0.46 1.23 0.23 1.07 0.63 0.19 1.16
57 methionol 0.03 0.15 N.D 2.09 0.23 N.D. 0.06 N.D. N.D. 5.27
58 valeric acid 5.11 6.93 16.36 5.15 5.19 4.00 5.65 8.16 6.84 6.30
59 2(5H)-furanone 0.27 0.33 0.25 0.13 0.12 0.26 0.29 0.40 0.26 0.19
60 2-acetyl-2-thiazoline N.D. 0.04 0.03 N.D. N.D. 0.02 N.D. 0.01 N.D. N.D.
61 δ-hexalactone 1.93 2.28 4.94 1.41 1.09 2.22 1.26 1.76 2.04 1.15
62 2-tridecanone 0.36 0.37 1.51 0.15 0.24 0.22 0.16 0.26 0.32 0.12
63 hexanoic acid 220.15 238.34 751.24 212.21 195.83 217.27 170.26 330.72 220.36 197.91
64 benzyl alcohol 0.27 0.70 0.62 0.23 0.20 0.53 1.31 2.34 0.18 1.36
65 dimethyl sulfone 81.43 81.35 119.39 43.88 31.26 20.46 43.10 36.92 54.08 35.08
66 phenethyl alcohol 0.14 0.40 0.53 0.07 0.16 0.03 0.26 0.08 0.17 0.53
67 heptanoic acid 1.63 2.79 6.24 0.97 0.73 1.04 1.85 3.11 2.01 1.50
68 benzothiazole 0.18 0.17 0.12 0.04 0.20 N.D. N.D. N.D. N.D. N.D.
69 maltol 0.42 2.90 7.69 0.30 0.15 0.23 0.37 0.71 0.96 1.03
70 δ-octalactone 1.40 1.54 4.19 0.92 0.58 1.71 0.46 1.48 1.46 0.67
71 phenol 0.03 0.06 0.06 0.02 0.05 0.02 0.02 0.03 0.02 0.02
72 2-pentadecanone 0.21 0.22 0.71 N.D. N.D. 0.24 0.10 0.25 0.14 0.12
73 2,5-dimethyl-4-hydroxy-3(2H)furanone 0.19 0.69 2.07 N.D. 0.09 0.09 0.06 0.08 1.79 0.10
74 ethyl myristate 0.10 1.19 0.54 0.11 0.07 0.03 0.04 0.58 0.18 0.03
75 octanoic acid 38.95 56.26 233.07 28.84 20.18 83.59 31.68 96.53 55.24 54.55
76 p-cresol 0.10 0.05 0.22 0.05 0.07 0.17 0.06 0.98 0.02 0.02
77 2-ethyl-4-hydroxy-5-methyl-3(2H)-furanone 0.35 0.26 2.08 N.D. N.D. 0.12 N.D. 4.33 1.51 N.D.
78 4-hydroxy-5-methyl-2(3H)furanone 0.41 7.57 8.22 N.D. 0.76 3.80 0.25 0.47 12.62 0.51
79 γ-decalactone 0.04 0.16 0.23 N.D. N.D. 0.05 0.11 0.08 0.12 0.02
80 nonanoic acid 1.80 2.52 5.39 0.75 0.71 1.04 0.74 2.21 1.79 0.84
81 lactic acid N.D. N.D. 33.27 N.D. N.D. N.D. N.D. N.D. N.D. N.D.
82 δ-decalactone 4.11 8.43 17.27 2.65 1.32 4.22 4.24 5.42 6.57 3.09
83 (E)-dec-8-en-5-olide N.D. N.D. N.D. N.D. N.D. 0.12 0.18 0.09 0.11 0.07
84 decanoic acid 26.57 32.73 153.09 12.42 6.65 45.42 13.31 55.39 31.36 27.74
85 glycerol N.D. 0.18 1.49 N.D. N.D. 0.11 0.22 0.31 0.17 0.18
86 δ-undecalactone 0.03 0.12 N.D. N.D. N.D. 0.07 0.08 30.10 0.11 0.04
87 9-decenoic acid 2.19 2.31 16.09 1.09 0.53 4.67 1.15 3.79 2.50 2.03
88 3-ethyl-4-methyl-2,5-pyrrolidinedione 0.19 0.82 0.81 0.24 N.D. 0.24 0.16 0.45 0.17 0.14
89 undecanoic acid 0.49 0.59 1.41 N.D. N.D. 0.37 N.D. 0.35 N.D. 0.17
90 γ-dodecalactone 0.45 1.16 2.04 0.24 0.11 0.63 0.45 0.62 1.20 0.23
91 (Z)-6-dodecen-4-olide N.D. 0.67 0.90 N.D. N.D. 0.07 0.29 0.11 0.67 0.05
92 benzoic acid 26.67 37.42 84.66 19.12 13.56 13.94 13.08 30.10 18.62 21.03
93 δ-dodecalactone 1.30 1.98 5.25 0.34 0.20 1.99 1.72 2.62 2.72 1.18
94 lauric acid 6.88 6.31 26.10 2.82 1.70 6.61 1.85 6.93 4.25 12.28
95 δ-tetradecalactone 0.62 0.85 1.98 0.14 N.D. 0.53 0.27 29.06 0.39 0.21
96 myristic acid 3.11 3.67 13.13 1.05 0.81 2.91 N.D. 2.75 0.82 1.05
97 tetradecenoic acid N.D. N.D. 1.66 N.D. N.D. N.D. N.D. N.D. N.D. N.D.
98 plamitic acid 7.33 8.05 18.45 3.40 1.81 N.D. N.D. N.D. N.D. N.D.
IS 3-heptanol 20.00 20.00 20.00 20.00 20.00 20.00 20.00 20.00 20.00 20.00
total 2216.33 1529.29 6781.06 1413.30 1696.94 1271.80 1385.61 1773.13 1578.13 1704.02
*  Unit; mg/ 100g, Determined by the ratio of the peak area of internal standard (added 20µg for cheese 100g). Response factor was 1.00.

**  ND; not detected

Table 3. Potent aroma components in the samples obtained using GC/O with Charm Analysis™
Number of detection order Descriptor Components Ri Charm Value
No.1 No.2 No.3 No.4 No.5 No.6 No.7 No.8 No.9 No.10
12 acidic/cheesy acetic acid 1430 18 0 131 20 103 0 83 14 27 78
20 butyric acid 1616 1264 1268 7504 1538 1921 1165 1284 1246 1562 2336
24 hexanoic acid 1840 77 98 354 73 25 81 45 106 86 50
38 decanoic acid* 2269 22 51 99 0 0 60 10 66 35 20
Total Acidic/cheesy 1381 1417 8088 1631 2049 1306 1422 1432 1710 2484
33 animal p-cresol 2069 33 15 59 0 25 51 14 338 0 0
41 3-methylindole* 2458 154 146 2109 48 343 501 91 328 96 149
Total Animal 187 161 2168 48 368 552 105 666 96 149
2 buttery diacetyl 1030 789 46 156 45 127 29 1477 48 22 48
Total Buttery 789 46 156 45 127 29 1477 48 22 48
1 cacao-like 2-/3-methylbutanal* 1027 21 16 0 893 244 63 344 48 42 166
Total Cacao-like 21 16 0 893 244 63 344 48 42 166
28 caramel unknown 12 1943 0 30 77 0 0 0 0 0 0 0
30 2,5-dimethyl-4-hydroxy-3(2H)furanone 2019 1084 6053 9094 28 650 533 621 865 8214 564
31,32 2-ethyl-4-hydroxy-5-methyl-3(2H)furanone and 4-hydroxy-5-methyl-3(2H)furanone* 2051 2482 2135 5697 20 376 1104 530 10349 3949 170
36 4,5-dimethyl-3-hydroxy-2(5H)furanone* 2173 111 194 1030 40 210 200 341 675 108 180
Total Caramel 3677 8412 15898 88 1236 1837 1492 11889 12271 914
9 cereal 2-acetyl-1-pyrroline* 1326 0 0 0 0 0 0 22 0 0 7
22 2-acetyl-2-thiazoline 1732 0 259 236 0 0 159 24 57 36 0
Total Cereal 0 259 236 0 0 159 46 57 36 7
7 fatty/metallic 1-octen-3-one* 1298 59 75 38 9 11 16 153 47 41 135
15 unknown7 1495 529 422 201 145 277 359 964 1341 863 867
16 (E)-2-nonenal* 1521 534 1125 633 2041 293 693 894 1956 371 1401
18 unknown8 1600 0 29 0 0 103 0 47 0 0 41
23 2,4-(E,E)-decadienal* 1791 14 111 0 0 0 0 152 9 40 15
29 trans-4,5-epoxy-(E)-2-decenal* 1985 22 66 63 12 208 48 776 240 102 22
Total Fatty/ metallic 1158 1828 935 2207 892 1116 2986 3593 1417 2481
4 floral/fruity unknown 1 1076 0 0 0 29 0 0 0 151 18 0
21 phenylacetaldehyde 1626 16 18 27 18 12 8 17 19 6 17
25 geraniol* 1846 39 0 45 37 8 8 42 44 41 20
26 unknown 10 1866 0 0 0 0 0 143 0 0 0 0
34 ethyl cinnamate * 2105 0 0 0 0 0 4038 23 0 0 0
37 o-aminoacetophenone* 2189 39 31 70 33 0 0 61 232 0 0
43 phenylacetic acid* 2548 0 11 105 27 0 0 23 84 0 0
Total Floral/ fruity 94 60 247 144 20 4197 166 530 65 37
5 green unknown2 1104 13 23 26 7 21 13 0 0 0 18
10 unknown4 1367 0 0 0 0 0 0 0 13 0 89
17 2,4-(E,E)-octadienal* 1572 26 0 0 44 10 0 0 51 9 94
6 unknown3 1145 0 0 0 0 0 0 0 47 0 72
Total Green 39 23 26 51 31 13 0 111 9 273
13 milky unknown6 1432 69 51 0 41 27 239 140 465 57 157
35 δ-decalactone 2160 126 242 274 42 14 150 132 174 227 44
39 γ-dodecalactone 2345 96 182 184 42 23 96 92 132 210 35
40 (Z)-6-dodecen-4-olide 2364 87 4887 6234 29 12 542 2253 566 4796 323
Total Milky 378 5362 6692 154 76 1027 2617 1337 5290 559
14 potato-like methional 1439 356 1236 295 7110 871 240 347 55 265 395
Total Potato-like 356 1236 295 7110 871 240 347 55 265 395
3 sulfur dimethyl disulfide* 1075 43 26 24 0 1948 0 60 0 0 14
8 2-methyl-3-furanthiol* 1311 0 0 36 0 0 0 0 0 0 0
11 unknown5 1424 13 5 0 0 18 0 0 0 0 0
19 unknown9 1605 12 14 29 0 0 0 0 0 0 0
Total Sulfur 68 45 89 0 1966 0 60 0 0 14
42 vanilla vanillin* 2532 84 15 61 52 0 0 47 0 0 0
Total Vanilla 84 15 61 52 0 0 47 0 0 0
27 woody unknown 11 1934 0 14 46 0 0 10 0 19 14 14
44 3-phenylpropionic acid* 2608 20 24 131 0 105 13 0 26 10 0
Total Woody 20 38 177 0 105 23 0 45 24 14
Total aroma intensity 8252 18918 35068 12423 7985 10562 11109 19811 21247 7541
*  aroma component only detected by GC/O.

Factor analysis for aroma descriptors    The factor analysis for 14 aroma descriptors found six characteristic aroma factors and the cumulative percent of factors was 91.7% (Table 4). The factor loading of ‘acidic/cheesy’ showed the highest value (0.95) among all the aroma descriptors in factor 1. The correlation coefficients were R = 0.85 between ‘acidic/ cheesy’ and ‘woody,’ and R = 0.88 between ‘animal’ and ‘woody’. Similarly, the factor loading of ‘milky’ showed the highest value (0.88) among all the aroma descriptors in factor 2. The correlation coefficient was R = 0.80 between ‘milky’ and ‘caramel.’ Factor 3 included ‘potato-like’ and ‘cacao-like,’ and the correlation coefficient was R = 0.90 between ‘potato-like’ and ‘cacao-like.’ Factor 3 was considered to be governed by sample No.4 due to the considerable amounts of methional (see Table 3). In the same manner, factor 5 was governed by sample No.5 and factor 6 by sample No.6, respectively.

Table 4. Contribution ratios of 14 aroma descriptors to six characteristic aroma factors
Aroma descriptors Factor 1 Factor 2 Factor 3 Factor 4 Factor 5 Factor 6
Acidic/cheesy 0.95
Animal 0.92
Woody 0.84
Milky 0.88
Caramel 0.71
Cereal 0.70
Green −0.58
Potato-like 0.99
Cacao-like 0.89
Buttery 0.90
Vanilla 0.84
Sulfur 0.82
Fatty/metallic −0.74
Floral/ fruity 0.96
Variance 3.20 2.54 2.14 1.81 1.75 1.40
Percent (%) 22.84 18.15 15.30 12.90 12.53 9.98
Cumulative percent (%) 22.84 40.98 56.28 69.18 81.71 91.69

Figure 1 shows the positioning map of the samples by factors 1 and 2, as these factors were considered to show a common characteristic aroma of Cheddar cheese. The palatability ranking and score of factors 1 and 2 are indicated in Table 5. Sample No.9 showed the highest value of consumer palatability, followed by sample No.8. On the other hand, samples No.4, No.5, and No.6 showed lower consumer palatability values than other samples. It should be noted that the aroma composition of these samples was governed by a single aroma component as described earlier (see Table 3). The results suggested that the aroma descriptors of ‘milky’, ‘caramel’, and ‘cereal’ had positive correlations with consumer palatability scores, while ‘acidic/cheesy’, ‘animal’, and ‘woody’ had negative correlations.

Fig. 1.

Positioning map of the samples by factors 1 and 2 based on CVs of 14 aroma descriptors (●; samples having common characteristic aromas, ○; samples dominated by a single aroma component).

Table 5. Scores of factors 1 and 2 as well as palatability
Sample Score Order of palatability
Factor 1 Factor 2 Palatability
No.1 −0.14 −0.61 5.93 6
No.2 −0.66 1.59 6.24 3
No.3 2.63 0.93 6.02 4
No.4 −0.17 −0.28 5.61 8
No.5 0.05 −1.11 5.31 10
No.6 −0.28 −0.15 5.51 9
No.7 −0.84 −0.03 5.97 5
No.8 −0.06 −0.07 6.27 2
No.9 −0.84 1.25 6.39 1
No.10 0.32 −1.53 5.75 7

Stepwise PLS-VIP identification of aroma components contributing to palatability    Prior to applying PLS to the charm values (CVs) of 43 aroma components and consumer palatability scores of the samples, three samples (No.4, No.5 and No.6) were excluded from the dataset because their aroma compositions were dominated by a single aroma component as mentioned earlier (see Table 3). The remaining seven samples were applied to further modeling processes of PLS-VIP and ANN.

The first PLS trial conducted for the CVs of 43 aroma components as explanatory variables and consumer palatability scores as objective variables was successfully predicted with high accuracy (R2 = 0.99). The stepwise PLS-VIP was applied to select the aroma components indicating a high contribution degree and order to palatability in the modeling. Figure 2 shows 28 aroma components selected as high contribution components in the first trial. Among all the aroma components, γ-dodecalactone (milky) showed the highest contribution to comprehensive palatability (VIP = 2.45). Aroma components highly contributing to the palatability scores of samples are shown in Fig. 2; thus, aroma components under the categories of ‘potato-like’ and ‘woody’ aroma were not included.

Fig. 2.

Contribution degree of 28 aroma components (threshold value; VIP = 0.80)

The trials were continued until all variables exceeded the selection threshold, and thus R2 = 0.72 was obtained at the 9th trial (Table 6). Numbers of the explanatory variables and coefficients of determination in the PLS model decreased with increasing number of trials.

Table 6. Stepwise selection of aroma components (threshold value; VIP = 0.80)
Trial time Objective Variable Explanatory Variable (Aroma component) R2 Potential factors
1th Trial Palatability (Sensory score) 43 0.99 2
2th 28 0.99 2
3th 21 0.79 2
4th 18 0.79 2
5th 16 0.76 2
6th 15 0.72 2
7th 14 0.67 2
8th 13 0.70 2
9th 12 0.72 3

The final results of stepwise PLS-VIP are presented in Fig. 3. Four aroma components such as δ-decalactone, 1-octen-3-one, acetic acid and 2,5-dimethyl- 4-hydroxy-3(2H)furanone were indicated as main aroma components of Cheddar cheese (Milo and Reineccius, 1997; Suriyaphan et al., 2001; Zehentbauer and Reineccius, 2002; Singh et al., 2003). All aroma components highly contributing to the palatability of Cheddar cheese were under the following 7 categories of aroma descriptors: ‘milky’; ‘green’; ‘fatty/metallic’; ‘acidic/cheesy’; ‘caramel’; ‘floral/fruity’; and ‘sulfur.’ It was also found that descriptors such as ‘milky’, ‘caramel’, and ‘floral/fruity’ had positive contributions to palatability scores, while ‘fatty/metallic’, ‘green’, ‘acidic/cheesy’, and ‘sulfur’ had negative contributions. In addition, the factor scores of ‘milky’ and ‘caramel’ increased with increasing palatability scores, while ‘acidic/cheesy’ increased with decreasing palatability scores (see Fig. 1 and Table 6).

Fig. 3.

Twelve aroma components obtained by stepwise PLS-VIP analysis contributing to the palatability of Cheddar cheese (threshold value; VIP = 0.80)

It should be noted that the sensory panelists in this study were all Japanese. As already mentioned, the Japanese prefer the relatively tasteless flavor and texture of processed cheeses compared to natural cheeses. These preferences are thought to be greatly influenced by the lengthy consumption of processed cheese. The results indicated the consumer preference of Japanese panelists for Cheddar cheese flavor.

Coupled modeling of PLS-VIP and ANN    PLS is one of the methods used to select variables on the assumption of data linearity, while ANN is the method used to flexibly model nonlinear relationships among the data. Aroma components highly contributing to palatability were specified by applying stepwise PLS-VIP to CVs of the GC/O data in the present study; however, the flavor of samples was not sufficiently characterized to permit its reproduction, by the mixture of pure compounds in a model cheese as mentioned earlier (Singh et al. 2003). This is probably due to the nonlinear relationships and mutual effects among the aroma components. Furthermore, the ‘component balance theory’ suggests that the Cheddar cheese flavor was produced by the correct balance and concentration of a wide range of sapid and aromatic compounds (Kosikowski and Mocquot, 1958), and thus it appeared difficult to model the Cheddar cheese aroma by simple linear modeling methods. Therefore, ANN, a nonlinear modeling method, was applied to the aroma components selected by PLS-VIP to model the relationship with comprehensive palatability scores.

The CVs of 12 aroma components identified by PLS-VIP and palatability scores were utilized as a dataset for basic ANN modeling to simulate their nonlinear relationships as well as the mutual effects contributing to palatability. The former data and the later scores were substituted into the input and output layers of ANN, respectively. Because the ANN method is a relatively flexible modeling method that can be easily over-fit to the applied data set, the over-fit penalty condition was set to 0.001 and the convergence criterion to 0.00001 to avoid over-fit modeling results.

Palatability was reasonably well predicted by the ANN model, indicating high accuracy values of 0.99 and 0.97 for R2c (calibration) and R2v (cross validation), as well as less than 0.01 and 0.05 for RMSEc and RMSEv, respectively. Figure 4(a) showed a positive effect of 2,5-dimethyl-4-hydroxy-3(2H)furanone (‘caramel’) on the local maximum value of palatability to provide a searching methodology for identifying the functional aroma component. The aroma components 2,5-dimethyl–4-hydroxyl–3(2H)furanone (‘caramel’) and γ-dodecalactone (‘milky’) indicated a positive contribution to palatability by PLS-VIP, providing convincing explanations of the palatability scores using ANN modeling.

Fig. 4.

Typical simulation results of ANN; (a) Seven samples were used for simulation. The CVs of 2,5-dimethyl-4-hydroxy-3(2H)furanone and γ-dodecalactone were used as aroma components contributing to palatability. (b) Ten samples were used for simulation. The CVs of dimethyl disulfide and γ-dodecalactone were used as aroma components contributing to palatability. (○; local maximum value of palatability)

The effects of eliminated aroma components of three samples (No.4, 5, and 6) on the local maximum value of palatability were investigated by adding their CVs into the input layer of the basic model. The model showed relatively lower accuracy values of 0.95 and 0.94 for R2c and R2v, as well as 0.05 and 0.03 for RMSEc and RMSEv, respectively. Figure 4(b) shows that the local maximum value and its value of palatability were markedly affected only by the CV for dimethyl disulfide of sample No.5 (see Table 3). As typically shown in the curvilinear characteristic between the local maximum value and origin in Fig. 4(b), the response curved surface was found to account for both nonlinear relationships and mutual effects among aroma components, as well as palatability. The optimal aroma balance of Cheddar cheese might be designable by using a coupled PLS-VIP and ANN method to simulate the relationships among aroma intensities and consumer palatability without making numerous prototypes.

The methodology of coupled stepwise PLS-VIP and ANN modeling based on the dataset of aroma components obtained from GC/MS as well as GC/O and sensory scores, provided a useful tool to evaluate the quality of Cheddar cheese, identifying the aroma components and their nonlinear relationships as well as mutual effects contributing to consumer palatability.

Personal attributes in the modified food kansei model affect the palatability of Cheddar cheese    Figure 5 shows the modified food kansei model (Ueda et al., 2008) summary of the main findings of the present study. The CVs of the aroma components were obtained by GC/O analysis as intrinsic attributes, and sensory evaluation was conducted by Japanese panelists on the basis of the perception route in the modified food kansei model. Concretely, the raw data of CVs was applied to PLS and stepwise PLS-VIP to select variables highly contributing to palatability scores. Furthermore, ANN was used to model nonlinear relationships among aroma components selected by PLS-VIP and palatability scores. In the present study, consumer evaluation of palatability was based on the cognition route in the modified food kansei model. The results of PLS-VIP analysis suggested that personal attributes such as food culture and dietary habits might influence the comprehensive palatability scores of Cheddar cheese, as shown in the modified food kansei model. It should be noted that personal attributes in the modified food kansei model should be clarified in implementing the consumer-oriented development of food products, especially of fermented food products.

Fig. 5.

Modified food kansei model (from Ueda et al., 2008)

Fermented food, including cheese, has various unique flavors that depend on the production area and basic ingredients. Thus, different people show dramatically different preferences for fermented food. The effect of personal attributes, as indicated in the modified food kansei model, will present as differences in the food cognition process of fermented food, based on the dietary habits and eating experiences of consumers.

The food kansei model was developed as a method to improve the quality of food products (Ikeda et al., 2004). To date, numerous studies have been performed based on this model, especially on the perception route of the model using a combination of various instrumental measurements and sensory evaluation. The modified model was proposed in consideration of personal attributes, as shown in Fig. 5 (Ueda et al., 2008). In previous studies regarding food kansei engineering, the dataset was preprocessed by principal component analysis or factor analysis. Then, the processed dataset, such as principal component scores or factor scores, was used as the variables for the model construction process. This resulted in difficulty in interpreting the principal components or factors derived from the analyses. In addition, since extrinsic food attributes in the cognition route of the food kansei model are difficult to quantify, previous studies have focused on the perception route (Ikeda et al., 2004; 2006; Shibata et al., 2008; Michishita et al., 2010). Although Akiyama et al. (2012) analyzed the effects of the diameter and color of straw on sensory cognition terms as a case study of the cognition route in the food kansei model, some difficulties remained in integrating the findings from the cognition route into the models constructed using the perception route data. The complexity of mutual effects among the information obtained from the routes of food perception and cognition makes it difficult to evaluate all processes.

The present study integrated the effects of personal attributes in the data analysis system as a factor in the consumer oriented product development and design method. Specifically, the main target consumers were set as Japanese housewives. Then, aroma components highly contributing to palatability scores were selected using PLS-VIP. In addition, ANN simulated the effects of aroma components on comprehensive palatability scores. A new methodology can be proposed to satisfy the preferences of specific consumers by using the data analysis procedure indicated in the present study. A further challenge is to develop a new method to merge the routes of human perception and cognition through additional case studies based on the food kansei model.

Conclusions

The instrumentally measured data (GC/MS and GC/O) for the aroma components of Cheddar cheese were correlated with the consumer palatability scores by PLS in order to develop a quality evaluation methodology for Cheddar cheese and to determine the contribution degree of aroma components in the aroma mixture.

Based on their order in the aroma composition, the aromas of “acidic/ cheesy”, “caramel” and “fatty/ metallic” were found to be the main aroma components of Cheddar cheese. PLS was conducted to construct predictive models of consumer palatability using CVs of the aroma components. Furthermore, stepwise PLS-VIP was used to narrow down the number of specific aroma components contributing to palatability. ANN modeling was used to demonstrate the nonlinear relationships among aroma components as well as palatability. Coupling stepwise PLS-VIP and ANN resulted in successful identification and ranking of aroma components contributing to the palatability of Cheddar cheese, and in modeling their nonlinear relationships.

In the present study, Japanese consumer panelists preferred mild flavors to characteristic cheese flavors. This tendency can be understood in the context of Japan's historical dietary culture regarding dairy products. Different people have different dietary cultures; therefore, consumer-oriented food product development requires consideration of personal attributes, shown in the modified food kansei model, especially in the case of fermented food products.

A further challenge is to perform validation tests of the model for predicting the comprehensive palatability scores of Cheddar cheese, which were evaluated by Japanese panelists in the present study. In addition, if similar experiments are performed for panelists in Euro-American nations having a long, rich dietary culture of cheese and other dairy products, entirely different results would be obtained. Focusing on personal attributes in the modified food kansei model would provide a useful cross-national perspective to compare food preferences, especially for fermented food products originating from different cultural spheres.

Acknowledgments    We gratefully acknowledge San-Ei Gen F.F.I., Inc. for their cooperation with the analyses of aroma components, and Morinaga Milk Industry, Co., Ltd. for supplying the Cheddar cheese samples.

References
 
© 2015 by Japanese Society for Food Science and Technology
feedback
 Top