Investigating the Variation of Volatile Compound Composition in Maotai-Flavoured Liquor During Its Multiple Fermentation Steps Using Statistical Methods

Maotai-fl avoured liquor, generally described as a highly complex-fl avoured, sweet and refreshing soy sauce aroma style alcoholic drink, is one of the most popular and representative liquors in China. The formation of the special fl avour of Maotai-fl avoured liquor can be largely att ributed to its unique and complicated production techniques. The process of Maotai-fl avoured liquor production diff ers from those of other liquors in many aspects including starter preparation, grain (mainly sorghum and wheat) piling and liquor distillation. Briefl y, the production of Maotai-fl avoured liquor consists in nine fermentation steps and the whole process lasts almost a year. Each fermentation step includes starter addition, piling (putting the mixture of cooked grains and starter powder on the ground, making it into a small hill, and then unde rgoing fermentation for 4–5 days), fermentation in a pit and dist illation. Aft er each fermentation step, the fermented mixture is distilled, the liquor is collected, and the fermented grains are used as the material for the next step. The liquor from the fi rst two fermentations, due to its coarse taste, is poured back on the piled mixture, while the liquor from the other seven fermentations is stored separately for further blending to form the fi nal product.


Introduction
Maotai-fl avoured liquor, generally described as a highly complex-fl avoured, sweet and refreshing soy sauce aroma style alcoholic drink, is one of the most popular and representative liquors in China.The formation of the special fl avour of Maotai-fl avoured liquor can be largely att ributed to its unique and complicated production techniques.The process of Maotai-fl avoured liquor production diff ers from those of other liquors in many aspects including starter preparation, grain (mainly sorghum and wheat) piling and liquor distillation.Briefl y, the production of Maotai-fl avoured liquor consists in nine fermentation steps and the whole process lasts almost a year.Each fermentation step includes starter addition, piling (put-ting the mixture of cooked grains and starter powder on the ground, making it into a small hill, and then unde rgoing fermentation for 4-5 days), fermentation in a pit and dist illation.Aft er each fermentation step, the fermented mixture is distilled, the liquor is collected, and the fermented grains are used as the material for the next step.The liquor from the fi rst two fermentations, due to its coarse taste, is poured back on the piled mixture, while the liquor from the other seven fermentations is stored separately for further blending to form the fi nal product.
Much eff ort has been made in recent years to fi nd the complicated fl avour composition of Maotai-fl avoured liquor, how it is formed and changes during the brewing process.The research includes microorganism composition analysis (1,2), the isolation and characterization of ISSN 1330-9862 scientifi c note doi: 10.17113/ft b.54.02.16.4582 functional strains (3), fl avour component determination (4,5), and the analysis of the relationships between the microorganisms and fl avour compounds (6,7).However, the utilization of multiple fermentation and distillation steps, a very special technique for Maotai-fl avoured liquor manufacturing, has not received enough academic att ention so far.
The fl avour of this Chinese liquor is rather complex and it is generally presented in the form of numerous gas chromatography (GC) or gas chromatography-mass spectrometry (GC-MS) data, for which statistical approach is a necessity to process and analyze the data.Cluster analysis (CA), discriminant analysis (DA), principal component analysis (PCA) and partial least square (PLS) regression have been widely adopted statistical methods in recent years.All these approaches have shown good performance in many cases of liquor fl avour research, such as spectral analysis (8), artifi cial nose (9), liquor discrimination and identifi cation (10)(11)(12).However, most of these methods are linear in nature, thus may not be capable of describing non-linear systems satisfactorily.As a promising alternative, artifi cial neural network (ANN) has a lot of advantages in parallel processing, classifi cation, learning and patt ern recognition.ANN has also been successfully used in researching the productions of wine and beer, such as prediction of process problems (13), sensory evaluation (14) and process optimization (15).As far as we know, ANN has seldom been applied in fl avour research of traditional Chinese liquor, especially M aotaifl avoured liquor, whose fl avour composition is regarded to be the most complicated among Chinese liquors.
The aim of this work, therefore, is to an alyze the variations of the fl avour composition of Maotai-fl avoured liquor during its multiple fermentation process with statistical approaches, and to provide useful information for bett er understanding of the formation of its fl avour style.

Liquor samples
The raw liquor samples were collected from nine distilleries of Langjiu Group Co., Ltd., Sichuan, PR China.Raw liquor was sampled aft er each of the seven fermentation and distillation steps.A total of 63 liquor samples were used for fl avour compound analysis.

Analytical methods
Gas chromatography (GC) analyses of liquor samples were performed on an Agilent 7890A gas chromatograph (Agilent Technologies Co. Ltd., Santa Clara, CA, USA) equipped with automatic sampler and fl ame ionization detector.Samples were analyzed on a CP-Wax 57 CB column (50 m×0.25 mm×0.2 μm).The injector, detector and column temperatures were set at 125, 120 and 90 °C, respectively.The carrier gas was N 2 .The fl ow rates of N 2 , H 2 and air were set at 20, 20 and 230 mL/min, respectively.A total of 68 fl avour compounds were determined by comparing their peak areas to those of the standards.All chemicals used in the analyses were of chromatographic grade.

Organoleptic evaluation
Organoleptic evaluation of liquor samples was conducted according to a literature method (16).All liquor samples were evaluated by ten tasters, and the average score of the fl avour and taste of each sample was calculated.

Data analysis
Student-Newman-Keuls test, correlation analysis and principal component analysis (PCA) were performed using SAS v. 8.1 soft ware (17).Cluster analysis, neural network model development and the calculation of each input neuron mean impact value (MIV) were carried out using MATLAB v. 7.1 soft ware (18).
Cluster analysis was used to group the liquor samples according to Euclidean distances between the samples based on their volatile compound compositions.PCA was applied to reduce the dimensionality of the original data matrix and allow the visualization of liquor samples with diff erent origins in a lower dimensional space.
Back-propagation neural network (BNN) models were established to predict fermentation steps and organoleptic evaluation scores of liquor samples based on their volatile compound compositions.The architecture of the neural network consisted of an input layer, a hidden layer and an output layer.The input nodes were the concentrations of the 68 volatile compounds of a liquor sample or corresponding principal components (PCs).The output is the fermentation step or the organoleptic evaluation score of the liquor sample.The number of the nodes in the hidden layer was selected according to the following equation: where n is the number of the nodes in the hidden layer, n 1 is the number of the input nodes, n 2 is the number of the output nodes, and a is a constant between 0-10.
The total dataset of 63 liquor samples was randomly split into two subdatasets, 48 samples for training and 15 samples for testing.The input variables for training and testing were standardized by using 'prestd' and 'trastd' functions, respectively, while the output variables were postprocessed by using a 'poststd' function in the neural network toolbox of MATLAB v. 7.1.For fermentation step prediction, the output variable was rounded to the nearest integral number.Bayesian regularization was adopted in training the neural network to avoid overtraining, and this was realized by using a 'trainbr' function in the neural network toolbox of MATLAB v. 7.1.Aft er BNN models with satisfactory predictive ability were established, MIVs were calculated to screen the most infl uential volatile compounds (19).

Variations of volatile compound composition of liquor samples from diff erent fermentation steps
The average concentrations of the volatile compounds in liquor samples from diff erent fermentation steps in the nine distilleries are listed in Table 1.It can be seen that n n n a    Correlation analysis suggests close relationships among some compounds (Table 2).These relationships can be part ly explained by t he sharing of common metabolic pathways or enzymes used for the fo rmation of diff erent compounds (acetaldehyde and acetal, n-propanal and n-propanol, 2-butanone and 2-butanol, for example).However, the close relationships among some other components (ethyl acetate and methanol, for example) are not fully understood and thus need further research in the future.

Cluster and principal component analyses of liquor samples after diff erent fermentation steps
Cluster analysis was p erformed to fi nd similarities among liquor samples aft er diff erent fermentation steps and distilleries based on their volatile compound compositions.The results show that except for several samples from the fi rst two fermentations, most samples from a same fermentation are clustered together (details not shown), suggesting that the fermentation step plays a more important role in forming the liquor style than the distillery where it is produced.
Comparing the samples from the same distillery, those aft er 3 to 7 fermentations are similar to each other, while those aft er the fi rst two rounds, especially the fi rst one, show mu ch diff erence (Fig. 1).This means that although there may be considerable diff erences in the be- Pearson's correlation coeffi cients among the components in the same rows are higher than 0.9 ginning, multiple fermentations and distillations lead to similar fl avour compound composition of the liquor aft er several steps.
In order to reduce the data dimensionality and visualize diff erent liquor samples in a lower dimensional space, PCA was conducted on the data matrix of 63 liquor samples×68 volatile compounds.The results reveal that the fi rst ten principal components (PCs) extracted are needed to account for 86 % of the total variance in the data matrix.The fi rst three PCs (PC1, PC2 and PC3), however, explain only 15.8, 14.1 and 9.4 % of the total variance, respectively.The three-dimensional plot of the PCA (Fig. 2) shows that liquor samples taken aft er fermentations 1 and 7 can be separated appropriately based on these three PCs, while other samples, especially those taken aft er fermentations 3-5, are very closely located.Samples from the same fermentation step from diff erent distilleries also failed to be separated satisfactorily from each other (details not shown).On the whole, the PCA here does not provide much insight for understanding the differences among the liquor samples.
Developing BNN models for predicting the n umber of fermentation steps and organoleptic evaluation score of liquor samples and variable screening BNN models were developed to predict the number of fermentation steps and organoleptic evaluation scores of the liquor samples based on their volatile compositions and PCs, respectively.Aft er a trial of the topological structure, it was found that BNN models with nine nodes in the hidden layer could provide satisfactory prediction when 68 volatile compound concentrations were used as inputs, while six nodes in the hidden layer were appropriate when ten PCs were used as inputs.Some representative predictions in the test are shown in Fig. 3.The accuracy of brewing round prediction in the test was between 80 and 100 % when 68 volatile compound concentrations were used as inputs, while 60-90 % of accuracy was obtained with ten PC inputs.For organoleptic evaluation score prediction, BNN model using 68 volatile compound concentrations as inputs also had bett er performance (R 2 value between 0.80 and 0.95) than those using the ten PCs as inputs (R 2 value between 0.70 and 0.90).This result was not entirely unexpected since the ten PCs account for only 86 % of the data variance and the linear PCA may lose some non-linear information of the investigated system.
As the BNN model with 68 volatile compound concentrations as inputs represents well the relationship between the volatile compositions and organoleptic evaluation results of the liquor samples, mean impact value (MIV) analysis was adopted to fi nd which volatile compounds play more important roles in forming the liquor fl avour style (Table 3).High MIVs of many alcohols and esters were observed.However, most of these MIVs are negative, suggesting that high concentration of these compounds may degrade the fl avour and taste of the liquor.Noticeable volatile compounds that showed positive and relatively high MIVs are ethyl lactate, furfural and several acids including valeric acid, heptanoic acid, isobutyric acid and nonanoic acid.This implies that these compounds contribute greatly to the formation of the fl avour of the liquor.The concentrations of almost all these compounds increased with the number of fermenations (Table 1), suggesting multiple fermentations a re vital in forming the liquor fl avour.
Major fl avour components in Maotai-fl avoured liquor have been discussed extensively in recent years but no consistent opinion has been obtained so far (20,21).Furfural (22), ac ids with high boiling point (23), and pyrazines ( 24) have all been suggested to be the major fl avour compounds in Maotai-fl avoured liquor.Our results show that ethyl lactate, furfural and some acids with high boiling points do play important roles in forming the liquor style.Pyrazines, however, seem to contribute less or even negatively according to their MIVs in Table 3.This is supported by a previous report where the concentration of pyrazines varied signifi cantly in diff erent Maotai-fl avoured liquor samples (20).However, as the analyses here are based exclusively on Maotai-fl avoured liquor samples, we cannot assert that the components with moderate or low MIVs are unimportant or even unnecessary in forming the liquor fl avour.Undoubtedly, further elucidation of major fl avour components in Maotai-fl avoured liquor requires more samples with diff erent fl avour characteristics.

Conclusion
The results of this research show that fermentation steps exert much more infl uence on the volatile compo sition of Maotai-fl avoured liquor than the distillery.Although there may be considerable diff erences in the vo latile com- The MIV of each compound is the average value of calculation results based on fi ve randomly selected trained BNN models position among the liquor samples at the be ginning, multiple fermentations and distillations ultimately lead to similar volatile composition of the liquor.Based on the statistical analyses, we suggest that ethyl lactate, furfural and some high-boiling-point acids make relatively high contribution in forming the special fl avour of Maotai-fl avoured liquor.

Fig. 1 .Fig. 2 . 7 Fig. 3 .
Fig. 1.A representative cluster analysis result of liquor samples from seven fermentation and distillation steps in a same distillery based on volatile compound composition

Table 1 .
Volatile compounds in the samples from diff erent fermentation steps of Maotai-fl avoured liquor production in nine distilleries

Table 1 .
-continued a Data are expressed as mean value±standard deviation (N=9).The same lett er in superscript in the same row denotes values that are not signifi cantly diff erent

Table 2 .
Close relationships among volatile compounds in Maotai-fl avoured liquor samples revealed by correlation analysis

Table 3 .
Mean impact value (MIV) analysis of the volatile compounds for the organoleptic evaluation score