1. Introduction
Ganoderma lucidum, a fungus commonly known as Lingzhi, has been used as a traditional functional food/medicine for centuries by rulers of the Chinese and Japanese dynasties to achieve enhanced vitality and longevity. These formulae take on exotic forms of special tea and mushroom concoction suitable for daily intake as supplements. Owing to the perceived benefits of these highly desirable medicinal properties localized within the spores (Lin 2001) and further amplified by profit focused producers, commercial demand for
The chemical structure of triterpenoids of
For this purpose, we developed a workflow involving the direct application of rapid FTIR and its feed-forward ANN model to perform classification of cracked spores of
Material and methods
Samples
120 samples comprised mainly of cultivated Ganoderma lucidum strain were analysed. These samples were received in bulk unpackaged powder form. To achieve model integrity, 90 samples comprising reference materials of cracked spores, uncracked spores and concentrate of
FTIR
Spectral measurements were carried out on a Shimadzu FTIR 8400S system equipped with a germanium coated KBr beam splitter, a Michelson type (30° incident angle) interferometer and a temperature controlled high sensitivity detector (DLATGS). Spectra were recorded under diffuse reflectance mode at a resolution of 4 cm-1 set to 128 scans. Spectra were acquired and processed using the as-supplied IRSolution program for microsoft windows. Absorbance peaks were then converted to second derivative using 15 convolution points (Munakata 1998). Calculation was performed by dividing the area of the absorption band of interest with a fixed frequency band located between 810 - 786 cm-1. This band was represented to be a silica-related absorption band (Bosch Reig et al. 2002) originating from the sample holder used to perform spectrum acquisition. In this paper, we considered using the area (relative) of each spectrum at pre-defined frequency bands to perform model construction. These frequency bands were represented by ν1 (1143 - 1037 cm-1), ν2 (1660 - 1560 cm-1), ν3 (1745 - 1716 cm-1) and ν4 (2845 - 2798 cm-1) respectively.
Statistics and data processing
Principal component analysis (PCA)
A commercially available Windows version of the JMP 9.0 (division of SAS Institute Inc, Cary, North Carolina, USA) was used to perform PCA, LDA and ANN analyses. For PCA analysis, the most prominent directions of the high-dimension data were identified. The dimensionality of the dataset was reduced by first applying a linear combination of the standardized original variables possessing the greatest possible variance, thereby creating the first principal component (PC 1). The second component (PC 2) was created based on the linear combination of the standardized original variables having the greatest possible variance and was uncorrelated with all previous defined components. By this principle, a reduced set of variables was achieved (Meloun et al. 1992). In this paper, unsupervised PCA was performed.
Linear discriminant analysis (LDA)
In DA, the classification variable is fixed and predicted by the continuous variables. JMP's implementation permits three types of DA(s) defined by linear, quadratic and mixed linear/quadratic. Briefly, LDA requires the Y variables to be normally distributed with the same variance and covariance but with different means for each group, while quadratic discriminate analysis requires the covariance to be different across the groups. In this paper, we considered applying LDA. To assess the model's classification robustness, 30 independent samples comprising cracked spores
Artificial neural network (ANN)
ANN analysis was used to assess the classification integrity predicted by the discriminant function, as well as to predict the amount of uncracked/cracked spores content by comparison with a known. Briefly, the TanH activation function was applied using one hidden layer. Values derived by applying PCA/LDA were fed into a hyperbolic tangent function (
Results
Differences in absorption spectra of Ganoderma lucidum samples
The absorption spectra of the homogenized samples were characterized by feature-rich frequency bands representative of major components of Ganoderma lucidum. These frequency bands are collectively defined by the fingerprints of polysaccharide, polysaccharide-peptide complex, β-glucans, lectins, organic germanium, adenosine, triterpenoids and nucleotides combined (Gao et al. 2004; Mizuno 1995; Liu 1999). Area-normalized spectra of cracked spores, uncracked spores and concentrate of
Area-normalised FTIR absorption spectra of a uncracked spores, b cracked spores and c concentrate of
Area-normalised FTIR absorption spectra of a uncracked spores, b cracked spores and c concentrate of
Principal component analysis (PCA), linear discriminant analysis (LDA)
In order to classify and quantify these variations, linear transformation method of PCA was first applied. 97.6% of the variance of the original dataset was explained by the first two PCs, as shown in the 2-D score plots of PCA results in Figure
2-D score plots of PCA results for a uncracked spores b cracked spores and c concentrate of
2-D score plots of PCA results for a uncracked spores b cracked spores and c concentrate of
Classification of uncracked spores (US), cracked spores (CS) and concentrate (C) of
Predicted group membership
US
CS
C
CS (Validation)
Total
Original
Count
US
30
0
0
0
30
CS
0
30
0
0
30
C
0
0
30
0
30
CS (Validation)
0
0
0
0
0
%
100.0
100.0
100.0
0
100.0b
Cross-validated
Count
US
30
0
0
0
30
CS
0
30
0
0
30
C
0
0
30
0
30
CS (Validation)
0
0
0
30a
30
%
100.0
100.0
100.0
100
100.0c
a. 30 independent samples of cracked spores of
b. 100.0% of original grouped cases were correctly classified.
c. 100.0% of cross-validated grouped cases were correctly classified.
Artificial neural network (ANN)
Coefficient of correlation (
Discussion
While it is possible to improve the first PC score further by reducing the variables (frequency bands), such approach raised some concerns within the framework of spectroscopy. Indeed, while PCA and LDA are useful tools suitably used to extract features that are focused on discriminating between classes via dimension reduction strategy, the error increment due to dimension reduction has to be without sacrificing the discriminative power of classifiers (Benediktsson and Sveinsson 1997). In this work, we did not observe such limitation. Rather, by shrinking the variables pool further, the advantage of utilizing the maximum chemical information provided by the absorption spectra will not be fully tapped. For this purpose, the values obtained for the PCs and canonical functions (LDA) were fed into an ANN model using 4 hidden nodes (33% random data holdback) to ascertain the classification outcome obtained when PCA and LDA were applied.
Within the framework of linear discriminant function, 3 distinct clusters representing cracked spores, uncracked spores and concentrate of
In brief, this work has examined the use of applying feed-forward ANN assisted by PCA/LDA analyses to discriminate cracked spores, uncracked spores and concentrate of
On hindsight, the successful marriage of spectroscopy and its statistical model perhaps lend light to the under-regulated functional food/TCM industry and its processes, towards achieving quality materials supply/control and quality products suitable for a safer public consumption objectives.
Completing interests
The authors declare that they have no competing interests.