Construction of a Prediction Model for Taste of Miso (Japanese Fermented Soybean Paste) Using Metabolic Profiling and Quantitative Descriptive Analyses

Abstract

Various types of miso are manufactured in Japan, which results in diverse flavors. However, the number of miso engineers who understand the variety of miso tastes has decreased due to decreases in the miso market. To solve this problem, we examined the potential to characterize various miso tastes using mainly low-molecular-weight hydrophilic component data by metabolic profiling. Analysis was performed by orthogonal partial least squares analysis, the relative intensity data for annotated compounds as explanatory variables and the data obtained by sensory evaluation as objective variables. The OPLS prediction model of taste had high accuracy and stability, and the components important for each taste were reasonable. From the results of model construction, we also found galactose correlated to umami. These results suggested that sensory evaluations of various types of miso may be in part replaced with instrumental analyses in the future.

Introduction

Various forms of brewed food have been manufactured in Japan since antiquity. Microorganisms (lactic acid bacteria, yeast, etc.) and enzymes produced by koji affect the taste and aroma of these brewed food products, as they produce various ingredients by the decomposition of raw materials and microbial metabolism. Miso is a representative brewed food made by inoculating Aspergillus oryzae into steamed cereals (rice, soybean, or barley), followed by storage in a high humidity environment with steamed soybeans, salt, and water for several months of ageing. Various types of miso are manufactured depending on the manufacturing region, such as rice koji-, soybean koji-, barley koji-, and mixed koji-based miso (a mixture of two types of koji). Furthermore, as conditions, such as the ingredient blending ratio, brewing period and temperature, and bacteria, vary among manufacturers, the metabolite composition, texture, and other properties also vary, resulting in diverse tastes, even for the same type of miso.

The number of miso manufacturers in Japan has been steadily decreasing. Therefore, the number of skilled engineers who understand the manufacturing and taste of miso is also decreasing. Furthermore, skill passing from experienced technicians with many years of experience and intuition to young engineers is difficult, and this is a major problem within the miso industry. Accordingly, it is necessary to develop a method for evaluating various miso tastes and for improving quality control.

The field of metabolomics, which enables comprehensive analyses of metabolites, has developed rapidly. In particular, food metabolomics contribute significantly to food safety and quality improvement (Castro-Puyana and Herrero, 2013; Hu and Xu, 2013; Wang et al., 2013). For example, metabolites in Japanese sake are correlated with its taste and aroma, and their importance has been determined by combining data from sensory evaluations, low-molecular-weight hydrophilic components, and aroma components (Mimura et al., 2014). In addition, metabolites correlated with the taste and aroma of soy sauce and their importance were determined from data for low-molecular-weight hydrophilic components, dipeptides, and sensory evaluations (Yamamoto et al., 2012, 2014).

Previous studies have investigated how various properties of miso affect its quality. For example, Kakezawa et al. analyzed miso in an annual competition and identified the relationship between physicochemical characteristics (glutamic acid, ethyl alcohol, water-resistant salt concentration, acidity, chromaticity, etc.) and high-quality miso (Kakezawa et al., 1983). A few groups have investigated differences in aroma components and the organic acid content between high- and low-ranked miso in an annual competition (Kato et al., 1983; Sugawara et al., 1992). A metabolic profiling study suggested that Amadori compounds contribute to aging in miso (Yoshida et al., 2009). Furthermore, the time course of the fermentation of doenjang, a traditional Korean seasoning similar to Japanese miso, has been analyzed by gas chromatography (GC)/mass spectrometry (MS), liquid chromatography/MS, and nuclear magnetic resonance, revealing metabolites that are correlated with quality (Lee et al., 2014; Namgung et al., 2010; Yang et al., 2009). However, comprehensive studies of the relationships between components and sensory characteristics for miso are lacking. Therefore, we have investigated the relationship between acylglycerols and quantitative descriptive analysis (QDA) (Ogawa et al., 2017). However, advanced equipment is required for the analysis of acylglycerols. Hence, it is more convenient to use components that can be measured with GCMS or a simple kit that can be used for quality analysis by miso manufacturer.

In this study, we constructed a model to predict the taste of miso (sweetness, salty taste, sourness, bitter taste, umami, and astringency) using various miso types in Japan and focusing mainly low-molecular-weight hydrophilic components by GCMS. Orthogonal projection to latent structures (OPLS) was used to construct the prediction model. The explanatory variables were data obtained by metabolic profiling which is mainly low-molecular-weight hydrophilic component; these components are thought to play an important role in taste. The objective variables were data obtained by QDA, a standard sensory evaluation method. Finally, we investigated the important components correlated with each taste property and examined the validity of the prediction model.

Material and Methods

Chemicals and materials    Chemical reagents, such as distilled water, methanol, chloroform, ribitol, and pyridine, were purchased from Wako Pure Chemical Industries, Ltd. (Osaka, Japan). Methoxyamine hydrochloride and alkane standard solution were purchased from Sigma-Aldrich (Milwaukee, WI, USA). MSTFA (N-methyl-N-(trimethylsilyl) trifluoroacetamide) was purchased from GL Science, Inc. (Tokyo, Japan).

Fourteen miso samples were purchased from stores and stored in a refrigerator at −30 °C. To analyze the ingredients in miso itself, most of the samples did not have additives. The samples, including koji, ingredients, and salinity, are summarized in Table 1. Samples were classified into four types: rice koji-based (which are either red or light-colored), soybean koji-based, barley koji-based, and those based on a mixture of rice and barley koji.

Table 1. List of samples

Number	Based Koji	Ingredients	Salinitya
1	rice	rice, soybean, salt	12.00%
2	rice	soybean, rice, salt	12.20%
3	rice	soybean, rice, salt	11.60%
4	rice	soybean, rice, salt, alcohol	12.00%
5	rice	soybean, rice, salt, alcohol	13.00%
6	rice	rice, soybean, salt, starch syrup, alcohol	4.90%
7	rice	rice, soybean, salt	10.00%
8	soybean	soybean, salt	10.30%
9	soybean	soybean, salt	11.40%
10	barley	barley, soybean, salt, alcohol	11.00%
11	barley	naked barley, soybean, salt	10.40%
12	mixed(rice and barley)	naked barley, soybean, rice, salt	10.50%
13	mixed(rice and barley)	rice, barley, soybean, salt	10.00%
14	mixed(rice and barley)	naked barley, soybean, rice, salt	9.50%

a:  indicated on the product label.

Sample Extraction    Miso is semisolid, with the potential for inhomogeneous distribution. Therefore, it is essential to homogenize miso. First, 5.0 g of each miso sample was placed into a tube, frozen with liquid-nitrogen, and crushed using a Multi-beads Shocker (Yasui-kikai, Kyoto, Japan). After returning to room temperature, each miso sample (10 mg) was placed in a 2-mL Eppendorf tube. To extract the hydrophilic components in miso by Bligh-Dyer extraction (Bligh and Dyer, 1959), a solvent (MeOH/H₂O/Chloroform, 5:2:2, v/v/v) mixed with ribitol (20 mg/mL) as an authentic standard was added to the 2-mL Eppendorf tube at 1 mL per tube. Then, centrifugal separation (9,390 G, 4 °C, 5 min) was performed for the removal of foreign bodies, 800 µL was transferred to another 2-mL Eppendorf tube, and 500 µL of distilled water was added, vortexed, and centrifuged (9,390 × g, 4 °C, 5 min). Then, 100 µL of the supernatant was placed in another 1.5-mL Eppendorf tube, centrifugal concentration was performed for 30 min, and the sample was freeze-dried overnight. All samples were analyzed in triplicate.

Sample derivatization    Oximation and trimethylsilylation were used for derivatization. First, 100 µL of methoxyamine hydrochloride (100 mL, 20 mg/mL in pyridine) was added to the dried sample and the mixture was incubated at 30 °C for 90 min (Thermomixer comfort, Eppendorf Co. Ltd., Tokyo, Japan). Then, 50 mL of MSTFA was added and the mixture was incubated at 37 °C for 30 min. The supernatant (100 µL) was added to a vial for GC injection.

GC/MS analysis    GC/MS was performed using the GCMS-QP2010 Ultra (Shimadzu, Tokyo, Japan). The InterCap 5MS/NP (0.25 mm × 30 m, 0.25 µm, GL Sciences, Tokyo, Japan) was prepared for the GC column. The AOC-20i/s (Shimadzu) auto-sampler was used for injection. Auto-tuning for the calibration of the MS detector was performed before the analysis.

Next, 1 µL of the above derivatized sample was injected into the GC/MS in a randomized order. The injection temperature was 230 °C in the split mode (50:1). The flow rate of the carrier gas He was 1.12 mL/min. The temperature of the column was held at 80 °C for 2 min, then increased 15 °C/min to 330 °C, and held for 6 min. The ion source temperature was 200 °C, and the interface temperature was 250 °C. The scan speed was 10000 u/s, the scan range was 85–500 m/z, and the EI voltage was 70 kV. In addition, Alkanmix (C10 to C38) (GL Sciences) was injected to calculate the retention index (RI).

Peak Identification    The MS spectral data obtained above were output in CDF format using GCMS Solution (Shimadzu). Baseline correction, alignment, and peak picking were performed and data were obtained in csv format using MetAlign (Lommen, 2009). Furthermore, using the semi-automatic identification software AIoutput (Tsugawa et al., 2011), annotation was performed based on the RI of the gas chromatograph and the electron ionization (EI) mass spectra. Finally, each intensity was divided by the intensity of ribitol, which was used as an internal standard.

Quantitative descriptive analysis    The QDA method and data are the same as those of a previous report (Ogawa et al., 2017). Number of panelists are 7. They first extracted attribute words for 14 miso samples. Secondly, data scaling (from 0 to 100) for the evaluation was adjusted between the panelists. Finally, quantitative data for all samples were acquired for three times. Results of the average data, standard deviation, ANOVA with two-way analysis of variance are added in Table S1.

Multivariate analysis    SIMCA-P13 (Umetrics, Umea, Sweden) was used for multivariate analysis. The intensities for each annotated compound divided by that of the internal standard (ribitol) were subjected to autoscaling (average = 0, variance = 1), and a principal component analysis (PCA) and OPLS analysis were performed. In the OPLS analysis, the compounds correlated with taste were determined from the relative intensity data for annotated compounds as explanatory variables and the data obtained by QDA as objective variables. R²Y (representing the predicted performance), Q²Y (the stability of the model), and VIP (Variable Importance in Projection; the magnitude of the influence of each component on each taste) were calculated.

Results and Discussion

Metabolic profiling results for miso    Based on the intensity data obtained by mass spectrometry, 36 compounds were annotated using AIoutput, including 15 amino acids, 5 organic acids, 10 sugars and sugar alcohols, and 6 others. The relative intensity, retention time (RT), RI, and m/z values are shown in Table S2 and in Fig. S1.

Fig. S1.

The relative quantitative data of each annotated compound of 14 samples of miso. The y-axis shows the relative intensity data of m/z for quantification of the annotated compounds divided by ribitol. The sample number is shown on x-axis.

Results of a PCA of the metabolic profile of miso    PCA was performed using the auto-scaled intensity data for 36 annotated compounds in 14 miso samples. The multivariate data were dimensionally compressed onto two axes explaining the greatest variance.

The score plot (Fig. 1 A) shows substantial clustering of soybean koji-based miso and other miso samples. In the loading plot (Fig. 1 B), glucose and trehalose, mainly found in rice koji and barley koji, respectively, were mostly in the minus direction of principal component (PC) 1. In contrast, the plus direction of PC1 was characterized by various amino acids (tyrosine, proline, threonine, glutamic acid, etc.) and organic acids (lactic acid, pyroglutamic acid, etc.). As soybeans and salt are the only ingredients in soybean koji-based miso, many amino acids in soybean miso are derived from soybean protein, and the amount of glucose is less than that in other types of miso. Among organic acids, glutamic acid is nonenzymatically changed to pyroglutamic acid; it promotes lactic acid fermentation during soybean koji production, since the fermentation period is long. This reflects the characteristics of these soybean koji-based miso types.

Fig. 1.

PCA score plot (A) and loading plot (B) for 36 annotated components in 14 miso samples. The score plot of the PCA can confirm similarity between samples, and the loading plot can visually show the metabolites present in the samples.

Next, to understand the difference between rice koji-based, barley koji-based, and mixed koji-based miso, PCA was performed without the soybean koji-based miso (No. 8 and No. 9). Clusters of miso were observed in the PC2-positive direction of the score plot, rice koji-based miso was observed in the minus direction of PC2 (Fig. S2 A), and the mixed koji-based miso was intermediate. In contrast, the axis showing the relative proportions of koji and soybean was in the PC1 direction of the loading plot because glucose was in the minus direction and various amino acids were in the positive direction. Mannitol and meso-erythritol, in the positive direction of PC2 (Fig. S2 B), are components of barley koji-based miso and are generally characterized by a larger proportion of koji than those of other misos. These components may also be used for markers to estimate the ratio of barley koji-miso in mixed koji-based miso.

Fig. S2.

PCA score plot A and loading plot B of sensory evaluation profiles in 12 miso samples. We excluded the soy bean koji-based miso (No.8 and No.9) in this PCA to understand the difference between rice koji-based, barley koji-based, and mixed koji-based miso. From the result of PCA score plot (A), we can see the clusters of barely koji based miso in the PC2 plus direction. On the other hand, the cluster of rice koji-based miso was observed in the minus direction of PC2. Then, the cluster of mixed koji-based miso was in between them.

From the result of PCA loading plot (B), the axis showing the proportion of koji and soybean was in the PC1 direction of the loading plot because the minus direction showed glucose and the plus direction showed various amino acids. Mannitol and meso-erythritol, which were located in the plus direction of the PC2 axis, are components produced by the barley koji-based miso and are generally characterized by a large proportion of koji compared to other miso.

PCA of QDA results    The results of ANOVA with two-way analysis of QDA show p < 0.01 for all attribute (Table S2), so we judged each attribute as significant enough to characterize PCA.

Next, PCA was performed using quantitative data for each term obtained by QDA. The score plot (Fig. 2 A) explicitly showed clusters of soybean-based koji miso and others; however, clear separation was not observed among rice koji-based, barley koji-based, and mixed koji-based miso. The results of the loading plot indicated that soybean koji-based miso has a characteristic bitter taste and aftertaste; astringency in taste; burnt, cocoa, and straw in the aroma; and straw and burnt qualities in the flavor (Fig. 2 B).

Fig. 2.

PCA score plot (A) and loading plot (B) of QDA profiles obtained for 14 miso samples.

Based on the QDA results, PCA was performed without soybean koji-based miso (No. 8 & No. 9), No. 6, and No. 13, because their sensory characteristics differed from those of other misos. Separation was observed (Fig. 3 A) in the score plot for rice koji-based miso, as shown in light shades (No. 1, 2, and 3) and in red (No.4 and 5). In addition, barley koji and mixed koji-based miso clusters were observed. The loading plot (Fig. 3 B) showed that the rice koji-based miso in light shades had a sweet and fruity flavor and aroma. The rice koji-based miso in red was characterized by a burnt aroma and flavor, roasted flavor, cocoa aroma, and bitterness in taste and aftertaste.

Fig. 3.

PCA score plot (A) and loading plot (B) of QDA profiles obtained for 10 miso samples, except for No 6,8,9 and 13

Summary of OPLS analysis of taste and annotated compounds    PLS is a popular analytical technique used in metabolomics, and multicollinearity can be avoided using latent variables (Triba et al., 2015). Furthermore, in OPLS analyses, prediction can be performed at a lower dimension by performing PLS to remove explanatory variables that are not related to objective variables. In this experiment, we aimed to construct the taste estimation model by OPLS analysis from the relative intensity data of annotated compounds, in other words the relative intensity data for annotated compounds as explanatory variables and the data obtained by QDA as objective variables. The results of OPLS analysis of taste are shown in Tables 2 and 3. Model performance can be evaluated by R²Y and Q²Y. R²Y shows the residuals of the objective variables that were not explained in the constructed model; values closer to 1 indicate that the model is more appropriate. Q²Y indicates the predicted performance of the model based on the data obtained by cross validation and is more stable as it approaches 1. More specifically, the cross validation is so called leave-one-out cross validations in this analysis. This is a statistical method to test whether it is possible to accurately predict the sample that has been removed in advance when one sample is removed from all the samples and a model is prepared using the remaining samples. The validation is irritated for all samples (in this time 14 times), then Q²Y is obtained.

Table 2. Summary of the results of OPLS model construction for taste

Attribute	No. of latent variable	R2X	R2Y	Q2Y
Sweetness	1+5+0	0.864	0.953	0.905
Aftertaste Sweetness	1+4+0	0.823	0.954	0.921
Saltiness	1+6+0	0.926	0.969	0.94
Aftertaste Saltiness	1+5+0	0.877	0.964	0.937
Sourness	1+4+0	0.824	0.975	0.957
Bitterness	1+4+0	0.825	0.968	0.942
Aftertaste Bitterness	1+4+0	0.828	0.977	0.96
Umami	1+6+0	0.925	0.963	0.93
Astringency	1+4+0	0.823	0.96	0.927
Rich Body	1+5+0	0.881	0.965	0.944

Table 3. VIP values of sensory terms for tastes from OPLS analysis (* indicates a negative value)

Sweetness VIP	Glucose 1.52	Malic acid 1.42*	Inositol 1.32*
Aftertaste Sweetness VIP	Glucose 1.51	Malic acid 1.31*	Phosphoric acid 1.31*
Saltiness VIP	Glucose 2.22*	Glycerol 1.93	Fructose 1.83*
Aftertaste Saltiness VIP	Glucose 2.15*	Glycerol 1.94	Fructose 1.91
Sourness VIP	Glucose 1.43*	Phosphoric acid 1.3	Proline 1.29
Bitterness VIP	Pyroglutamic acid 1.34	Proline 1.31	Valine 1.3
Aftertaste Bitterness VIP	Pyroglutamic acid 1.33	Phosphoric acid 1.33	Proline 1.3
Umami VIP	Galactose 3.17	Arabinose 1.8	Melibiose 1.79
Astringency VIP	Pyroglutamic acid 1.35	Threonine 1.34	Serine 1.33
Rich Body VIP	Galactose 1.76	Threonine 1.31*	Serine 1.27*

On OPLS analysis, R²Y and Q²Y were very close to 1, suggesting that a good prediction model for each taste has been constructed. Even using various kinds of miso in Japan, this stability of model (Q²Y) is high. Therefore, it seems that the number of samples to make prediction a model for taste is sufficient.

Next, we investigated the validity of annotated compounds with high VIP values obtained from the results summarized in Table 3. Glucose was positively correlated with sweetness and negatively correlated with a salty taste, both of which are expected results; glucose seems to directly contribute to sweetness and salty taste. However, glycerol correlated with salty taste, despite the fact that glycerol itself is sweet. Glycerol in miso is produced in multiple ways. First, triacylglycerol (TAG) in soybeans is degraded by lipases, yielding glycerol and fatty acids. Second, yeasts (including Zygosaccharomyces rouxii, salt-tolerant yeast in miso) produce glycerol to adjust the intracellular osmotic pressure as their sugar and salt concentrations increase. Considering rice koji-based miso as an example, if the amount of soybean increases in proportion to the amounts of raw materials used for miso production, the amount of TAG would increase. Therefore, glycerol decomposed by TAG would also increase. Increasing the ratio of quantity of soybeans to miso, the amount of rice koji to miso, which is the source of glucose, would decrease. As a result, it is assumed that glycerol is positively correlated to saltiness. Moreover, it is thought that yeast accumulates glycerol in response to the high concentration of salt in miso; therefore, it is considered that saltiness and glycerol are correlated. For sourness, glucose is negatively correlated from the QDA data of sourness, the highest values were seen in No.8 and No.9, which are in the soybean-miso group. On the other hand, the lowest value was seen in No.6. Soybean miso (No.8 and No.9) generally does not have glucose, as compared to other types of miso, because they do not use rice-koji or barley-koji, and No.6 miso is abundant in glucose. This situation seems to suggest that glucose and sourness are correlated. Inconsistent with expectations, monosaccharide galactose showed the highest VIP value for umami. Basically, as galactose is mainly derived from soybeans in miso (Hondo and Mochizuki, 1979), it might be interpreted that umami will increase as the amount of soybean increases. Thus, it is suggested that galactose may become a biomarker of umami. Generally, sodium glutamate is known as the umami component of miso and soy sauce; therefore, some manufacturers add it to increase the umami flavor. Accordingly, we expected glutamic acid to show the highest VIP value. However, we speculated that differences in umami could not be effectively judged by directly licking miso during the sensory evaluation owing to its strength. Pyroglutamic acid, which is inherently tasteless, showed high VIP values for bitterness and astringency. Generally, a longer fermentation period and higher aging temperature are related to greater glutamate conversion to pyroglutamic acid by a nonenzymatic reaction. Therefore, pyroglutamic acid itself is indirectly correlated to bitterness and astringency. Thus, changes in other ingredients due to the aging period and temperature may contribute to bitterness and astringency. Although proline itself is sweet, proline is correlated to bitterness in this experiment. Reports have shown that a proline-bearing peptide has the greatest bitterness (Kubo et al., 1988). Therefore, it is possible that miso with high proline content may be bitter because peptides containing proline are relatively abundant. Metabolic profiling of peptides is required to confirm this.

In summary, metabolic profiling was used to evaluate a wide variety of misos in Japan. The components and sensory terms for each category of miso were investigated by PCA and the characteristics and attributes of various types of miso had revealed for types of miso. Using metabolic profiling and QDA data, we successfully constructed a good prediction model for taste based on mainly 36 low-molecular-weight components, which means the model is not only appropriate (R²Y is close to 1), but also stable (Q²Y is close to 1) in these 14 miso samples. We also found compounds that are correlated with miso taste from the OPLS analysis. In addition, some of components with high VIP values for each taste correlated to the prediction of taste, thus supporting the validity of the prediction model. In particular, we found that galactose may become a biomarker of umami. These results imply the potential to evaluate miso taste by instrumental analyses. Sensory evaluations of various types of miso can be replaced with data obtained from instrumental analyses in the future.

Acknowledgements    This research was supported in part by “The special scheme to deploy highly advanced technology for agriculture, forestry and fisheries” funded by the Ministry of Agriculture, Forestry and Fisheries.

The study represents a portion of the dissertation submitted by Takahiro Ogawa to Osaka University in partial fulfillment of the requirement for his PhD.

References

•  Bligh,  E. and  Dyer,  W.J. (1959). A rapid method of total lipid extraction and purification. Can. J. Biochem., 37, 911-917.
•  Castro-Puyana,  M. and  Herrero,  M. (2013). Metabolomics approaches based on mass spectrometry for food safety, quality and traceability. TrAC - Trends Anal. Chem., 52, 74-87.
•  Hondo,  S. and  Mochizuki,  T. (1979). Polysaccharides in Soybean Steamed Waste Water and Miso. J. Jpn. Soc. Food Sci. Technol. (Nippon Shokuhin Kogyo Gakkaishi), 26, 461-468.
•  Hu,  C. and  Xu,  G. (2013). Mass-spectrometry-based metabolomics analysis for foodomics. TrAC - Trends Anal. Chem., 52, 36-46.
•  Kakezawa,  M.,  Ishima,  T., and  Ebine,  H. (1983). On Light Yellow Salty Rice Miso, Soybean Miso, Barley Miso and White Miso Relation Between Sensory Evaluation and the Results of Physical and Chemical Analyses on Miso by Component Analysis Method (Part I). J. Brew. Soc. Japan, 78, 713-720 (In Japanese).
•  Kato,  M.,  Kikuchi,  K.,  Mori,  Y.,  Tabei,  H., and  Takami,  I. (1983). Carboxylic Acids Contents and Composition in Salty Rice Miso. J. Jpn. Soc. Food Sci. Technol. (Nippon Shokuhin Kogyo Gakkaishi), 30, 99-107 (In Japanese).
•  Kubo,  T.,  Chino,  M., and  Fukui,  H. (1988). Taste of Proline-containing The taste of peptides composedof proline sequently. Agric. Biol. Chem., 52, 95-98.
•  Lee,  S.Y.,  Lee,  S.,  Lee,  S.,  Oh,  J.Y.,  Jeon,  E.J.,  Ryu,  H.S., and  Lee,  C.H. (2014). Primary and secondary metabolite profiling of doenjang, a fermented soybean paste during industrial processing. Food Chem., 165, 157-166.
•  Lommen,  A. (2009). Metalign: Interface-driven, versatile metabolomics tool for hyphenated full-scan mass spectrometry data preprocessing. Anal. Chem., 81, 3079-3086.
•  Mimura,  N.,  Isogai,  A.,  Iwashita,  K.,  Bamba,  T., and  Fukusaki,  E. (2014). Gas chromatography/mass spectrometry based component profiling and quality prediction for Japanese sake. J. Biosci. Bioeng., 118, 406-414.
•  Namgung,  H.J.,  Park,  H.J.,  Cho,  I.H.,  Choi,  H.K.,  Kwon,  D.Y.,  Shim,  S.M., and  Kim,  Y.S. (2010). Metabolite profiling of doenjang, fermented soybean paste, during fermentation. J. Sci. Food Agric., 90, 1926-1935.
•  Ogawa,  T.,  Izumi,  Y.,  Kusumoto,  K.,  Fukusaki,  E., and  Bamba,  T. (2017). Wide target analysis of acylglycerols in miso (Japanese fermented soybean paste) by supercritical fluid chromatography coupled with triple quadrupole mass spectrometry and the analysis of the correlation between taste and both acylglycerols and free fatty. Rapid Commun. Mass Spectrom., 31, 928-936.
•  Sugawara,  E.,  Saiga,  S., and  Kobayashi,  A. (1992). Relationships between Aroma Components and Sensory Evaluation of Miso. J. Jpn. Soc. Food Sci. Technol. (Nippon Shokuhin Kogyo Gakkaishi), 39, 1098-1104 (In Japanese).
•  Triba,  M.N.,  Moyec,  L.,  Le Amathieu,  R.,  Goossens,  C.,  Bouchemal,  N.,  Nahon,  P.,  Rutledgee,  D.N., and  Philippe  Savarin (2015). PLS/OPLS models in metabolomics: Impact of permutation of dataset rows on the K-fold cross-validation quality parameters. Mol. Biosyst., 13-19.
•  Tsugawa,  H.,  Tsujimoto,  Y.,  Arita,  M.,  Bamba,  T., and  Fukusaki,  E. (2011). GC/MS based metabolomics: development of a data mining system for metabolite identification by using soft independent modeling of class analogy (SIMCA). BMC Bioinformatics, 12, 131.
•  Wang,  X.,  Wang,  S., and  Cai,  Z. (2013). The latest developments and applications of mass spectrometry in food-safety and quality analysis. TrAC - Trends Anal. Chem., 52, 170-185.
•  Yamamoto,  S.,  Bamba,  T.,  Sano,  A.,  Kodama,  Y.,  Imamura,  M.,  Obata,  A., and  Fukusaki,  E. (2012). Metabolite profiling of soy sauce using gas chromatography with time-of-flight mass spectrometry and analysis of correlation with quantitative descriptive analysis. J. Biosci. Bioeng., 114, 170-175.
•  Yamamoto,  S.,  Shiga,  K.,  Kodama,  Y.,  Imamura,  M.,  Uchida,  R.,  Obata,  A.,  Bamba,  T., and  Fukusaki,  E. (2014). Analysis of the correlation between dipeptides and taste differences among soy sauces by using metabolomics-based component profiling. J. Biosci. Bioeng., 118, 56-63.
•  Yang,  S.-O.,  Kim,  M.-S.,  Liu,  K.-H.,  Auh,  J.-H.,  Kim,  Y.-S.,  Kwon,  D.Y., and  Choi,  H.-K. (2009). Classification of fermented soybean paste during fermentation by 1H nuclear magnetic resonance spectroscopy and principal component analysis. Biosci. Biotechnol. Biochem., 73, 502-507.
•  Yoshida,  H.,  Yamazaki,  J.,  Ozawa,  S.,  Mizukoshi,  T., and  Miyano,  H. (2009). Advantage of LC-MS Metabolomics Methodology Targeting Hydrophilic Compounds in the Studies of Fermented Food Samples. J. Agric. Food Chem., 57, 1119-1126.