Research Article Received: 16 August 2013,
Revised: 28 November 2013,
Accepted: 30 November 2013
Published online in Wiley Online Library
(wileyonlinelibrary.com) DOI 10.1002/pca.2502
Thin-layer Chromatographic Identification of Chinese Propolis Using Chemometric Fingerprinting Tie-xin Tang,a Wei-yan Guo,a Ye Xu,a Si-ming Zhang,a Xin-jun Xu,a,b Dong-mei Wang,a Zhi-min Zhao,a,b Long-ping Zhua,b and De-po Yanga,b* ABSTRACT: Introduction – Poplar tree gum has a similar chemical composition and appearance to Chinese propolis (bee glue) and has been widely used as a counterfeit propolis because Chinese propolis is typically the poplar-type propolis, the chemical composition of which is determined mainly by the resin of poplar trees. The discrimination of Chinese propolis from poplar tree gum is a challenging task. Objective – To develop a rapid thin-layer chromatographic (TLC) identification method using chemometric fingerprinting to discriminate Chinese propolis from poplar tree gum. Methods – A new TLC method using a combination of ammonia and hydrogen peroxide vapours as the visualisation reagent was developed to characterise the chemical profile of Chinese propolis. Three separate people performed TLC on eight Chinese propolis samples and three poplar tree gum samples of varying origins. Five chemometric methods, including similarity analysis, hierarchical clustering, k-means clustering, neural network and support vector machine, were compared for use in classifying the samples based on their densitograms obtained from the TLC chromatograms via image analysis. Results – Hierarchical clustering, neural network and support vector machine analyses achieved a correct classification rate of 100% in classifying the samples. A strategy for TLC identification of Chinese propolis using chemometric fingerprinting was proposed and it provided accurate sample classification. Conclusion – The study has shown that the TLC identification method using chemometric fingerprinting is a rapid, low-cost method for the discrimination of Chinese propolis from poplar tree gum and may be used for the quality control of Chinese propolis. Copyright © 2014 John Wiley & Sons, Ltd. Keywords: chemometrics; fingerprinting; identification; image analysis; pattern recognition; thin-layer chromatography; Chinese propolis
Introduction Propolis is produced by worker bees, which collect resin from plants, mix it with their salivary secretions, and finally add beeswax to produce the propolis (bee glue). Propolis has been marketed and used extensively in health foods, cosmetics and traditional medicines for many years. It exhibits numerous beneficial bioactivities, including anti-inflammatory, anti-oxidant, anti-bacterial, anti-viral, anti-fungal, immunomodulatory and anti-tumour effects (Stefano and Francesco, 2002; Salatino et al., 2011; Sforcin and Bankova, 2011). In Chinese pharmacopeia, propolis is listed as a drug for treating hyperlipidaemia, diabetes, etc., in the volume on traditional Chinese medicine (Committee of National Pharmacopeia, 2010). Plant resins are the primary material used by bees to produce propolis and they determine its chemical composition. The majority of Chinese propolis originates from the Populus species and its hybrids. Chinese propolis is typically classified as poplartype (Wu et al., 2008). Its major bioactive components are polyphenols such as flavonoids and caffeic-acid derivatives (Gardana et al., 2007; Medana et al., 2008; Salatino et al., 2011). Poplar tree gum, the extract of poplar buds, has a similar chemical composition and appearance to Chinese propolis and has been widely used as a counterfeit (Wu et al., 2008; Zhang et al., 2011). China is the world’s second largest producer of propolis (Paviani et al., 2010), and it
Phytochem. Anal. 2014
is extremely important to discriminate Chinese propolis from poplar tree gum. Although numerous studies have addressed the discrimination and fingerprinting of propolis samples (Gardana et al., 2007; Medana et al., 2008; Zhou et al., 2008; Nagaraja et al., 2010; Sârbu and Moţ, 2011; Cheng et al., 2013), few have specifically tackled the discrimination of Chinese propolis from poplar tree gum. A method employing Fourier transform infrared spectroscopy combined with two-dimensional infrared correlation analysis was developed to distinguish the extracts of Chinese propolis from the extracts of poplar buds (Wu et al., 2008). In another approach, HPLC of salicin was used to screen for counterfeit Chinese propolis (Zhang et al., 2011). Unfortunately, these methods require the use of expensive instrumentation.
* Correspondence to: De-po Yang, School of Pharmaceutical Sciences, Sun Yat-sen University, Guangzhou, China 510006. E-mail: [email protected] a
School of Pharmaceutical Sciences, Sun Yat-sen University, Guangzhou, China, 510006
b
Guangdong Technology Research Center for Advanced Chinese Medicine, Guangzhou, China, 510006
Copyright © 2014 John Wiley & Sons, Ltd.
T.-x. Tang et al. In contrast, TLC is simple, flexible, accessible and affordable. It is the most commonly used identification method for herbal medicines in the pharmacopeias of countries all over the world. Thin-layer chromatographic fingerprinting is most often conducted in a subjective manner using manually determined peak differences, and the application of chemometrics to TLC classification is still under development. However, TLC chromatograms can be captured using a digital camera, and image and chemometric analyses can be performed on a personal computer. The combination of TLC with image analysis and chemometrics can provide a fascinating, rapid and low-cost analytical tool for the identification of natural products. Several studies have reported the application of chemometrics to TLC using image analysis for data acquisition. Principal component analysis (PCA), partial-least-squares discriminant analysis (PLS-DA) and orthogonal PLS-DA were used for a metabolomics study of the HPTLC fingerprints of four medicinal herbs (Ogegbo et al., 2012). Thin-layer chromatography was combined with image analysis and chemometric analyses, including similarity analysis (SA), artificial neural network (ANN, otherwise known as neural network, NN), and k-nearest neighbours (k-NNs), to classify samples of the genus Bupleuri radix (Tian et al., 2009). Principal components analysis and hierarchical clustering (HC) analysis were used with thermostated microthin-layer chromatography for the rapid screening and classification of spirulina from pharmaceutical formulations and food samples (Zarzycki et al., 2011). Thin-layer chromatography was combined with image analysis and hierarchical fuzzy clustering for the successful ecosystem discrimination and fingerprinting of Romanian propolis (Sârbu and Moţ, 2011). In the above cases, although the sample separation on the TLC plate was not good, the chemical profile of the samples resulted in the formation of unique features on the TLC chromatograms. Using chemometric methods, the samples could be classified correctly based on the TLC results, suggesting that TLC combined with image analysis and chemometrics can be used for the identification of Chinese propolis. In this study, a new TLC method for the identification of Chinese propolis was developed. The chemometric methods SA, HC, k-means clustering (KMC), NN and support vector machine (SVM) were compared for use in classifying the samples based on their TLC densitograms. Furthermore, a strategy to discriminate between Chinese propolis and poplar tree gum samples is proposed.
Experimental Materials and reagents Chinese propolis samples and poplar-tree gum samples were collected from beekeepers or supplied by commercial companies in China. The details of the samples studied are provided in Table 1. Reference standards for quercetin (S1), kaempferol (S2), galangin (S3) and chrysin (S4) were purchased from the National Institute for the Control of Pharmaceutical and Biological Products (Beijing, China). Methanol, toluene, ethyl acetate, formic acid, ammonia aqueous solution (25%) and hydrogen peroxide aqueous solution (30%), from the Guangzhou Chemical Reagent Factory (Guangzhou, China), were of analytical grade. Silica gel 60 pre-coated TLC plates (100 × 200 mm) were purchased from Qingdao Haiyang Chemical Co., Ltd (Qingdao, China). The samples and reference standards were weighed using an AG285 analytical balance (Mettler Toledo Instruments AG, Zurich, Switzerland). Ultrasonic extraction was accomplished using an SB25-12D ultrasonic cleaner (Ningbo Scientz Bio-tech Co., Ltd, Ningbo, China). The solutions were applied to the TLC plate using a microlitre syringe (Shanghai GaoKe Industrial and Trading Co., Ltd, Shanghai, China). The plates were developed in a 200 × 200 mm (length × height) twin-trough glass chamber (Shanghai Xinyi Instrument Co., Ltd, Shanghai, China). The TLC images were captured using an A700 digital camera (Cannon Inc., Tokyo, Japan).
Extraction and thin-layer chromatography Accurately weighed 0.5 g samples were extracted with 20 mL of methanol for 20 min by ultrasonic extraction. The extract solution was filtered and used as the sample solution. Accurately weighed quercetin, kaempferol, galangin and chrysin were dissolved in appropriate volumes of methanol to produce solutions of 0.254, 0.258, 0.255 and 0.251 mg/mL, respectively. The sample solutions and standard solutions were applied in spots on a TLC plate. A mixture of toluene:ethyl acetate:formic acid (10:4:0.5) was used as the mobile phase. A twin-trough chamber was presaturated with the mobile phase for 15 min, and the plate was developed in the chamber until the solvent front had migrated approximately 7 cm from the origin. The plate was then removed from the chamber, dried in an oven at 105 °C for 10 min, and placed on a rack in another twin-trough chamber. An ammonia aqueous solution of 10 mL was added to one trough of the second chamber, and a hydrogen peroxide aqueous solution of 10 mL was added to the other trough. The second chamber was covered with a lid, and the plate was treated with the ammonia and hydrogen peroxide vapours until the colour of the spots on the plate had stabilised. Finally, the plate was removed and photographed under natural light. Three separate people performed the TLC analysis in order to obtain more data to train and test the discrimination models in chemometric fingerprinting.
Table 1. Origins of the Chinese propolis and poplar-tree gum samples Samples Chinese propolis Chinese propolis Chinese propolis Chinese propolis Chinese propolis Chinese propolis Chinese propolis Chinese propolis Poplar tree gum Poplar tree gum Poplar tree gum
Sample codes
Suppliers
T1 T2 T3 T4 T5 T6 T7 T8 F1 F2 F3
Xiangyun bee yard, Xuancheng, Anhui Zhuxuefeng bee yard, Qingdao, Shandong Lengjiao bee-related products store, Meishan, Sichuan Purui bee-related products company, Xi’an, Shangxi Yifengtang bee-related products store, Xi’an, Shangxi Luohandong bee yard, Dongda county, Xi’an, Shangxi Dahua bee-related products company, Zhengzhou, Henan (collected from Shandong) Dahua bee-related products company, Zhengzhou, Henan (collected from Anhui) Xiemei bee-related products company, Shanghai Xilian bee-related products company, Lanzhou, Gansu Dahua bee-related products company, Zhengzhou, Henan
wileyonlinelibrary.com/journal/pca
Copyright © 2014 John Wiley & Sons, Ltd.
Phytochem. Anal. 2014
Chemometric TLC Identification of Chinese Propolis Image analysis Images of the TLC chromatograms, saved in JPEG format, were analysed using the self-programmed software developed by Delphi Xe2 IDE (Embarcadero, San Francisco, CA, USA). First, the image was pre-treated by manually rotating to horizontally align the initial sample spots. The image was then manually cropped from the initial sample spots to the solvent front to eliminate the irrelevant regions. Subsequently, the image was automatically converted to greyscale and scaled to a height of 256 pixels. Second, the image was virtually scanned and automatically segmented into lanes corresponding to the sample migration paths. By summing the grey values of the pixels in each row at the same migration distance in the lane, a one-dimensional array was automatically obtained. A densitogram was then produced by plotting the line graph on an axis with the data of the one-dimensional array. Finally, the densitograms were manually offset by fixing the Rf of common compounds to the same value (Sârbu and Moţ, 2011), and the numeric variables of each densitogram were automatically scaled in the range from 0 to 255. Thin-layer chromatographic greyscale chromatograms reconstructed from the densitograms, were created by applying the densitogram data as the greyscale value to the pixels of a line, and repeating the line 28 times.
Chemometric analysis One densitogram (256 variables) was used as a dataset for chemometric analysis. A total of 33 densitograms (11 samples × 3) constituted the whole datasets. Similarity, HC, KMC and NN analyses were performed by SPSS (Version 21, IBM Corporation, Armonk, New York, USA). Support vector machine analysis was performed by LIBSVM (Version 3.16) (Chang and Lin, 2011). The options were optimised to obtain the best results.
Results and discussion Thin-layer chromatography of the samples and its image analysis A toluene:ethyl acetate:formic acid (30:12:5, v/v/v) mixture was used as the mobile phase for the TLC of Romanian propolis samples according to the previous method (Sârbu and Moţ, 2011). However, this mixture was not suitable for the TLC of Chinese propolis using silica gel 60 pre-coated plates. The solvent ratio was therefore adjusted, and a toluene:ethyl acetate:formic acid ratio of 10:4:0.5 was found to provide improved separation. In the previous method (Sârbu and Moţ, 2011) the TLC of Romania propolis samples was visualised by spraying a methanolic solution of 0.2% diphenylboryloxyethylamine and 4% polyethyleneglycol on the plate and capturing a fluorescent image under UV 366 nm light using a UV viewer cabinet. To circumvent the need to use fluorescent imaging equipment, a new visualisation method was developed using ammonia and
hydrogen peroxide vapours. The hydrogen peroxide was used as the oxidant, and the ammonia vapour ionised the polyphenols, increasing their susceptibility to oxidation. The TLC chromatograms of the samples, performed separately by three individuals, are displayed in Fig. 1. The polyphenols were oxidised by the ammonia and hydrogen peroxide vapours and appear different in colour. There were five main spots in the chromatograms of the Chinese propolis samples. Four of the spots corresponded in Rf values to those in the chromatograms of S1, S2, S3 and S4. The chromatograms of the three poplar-tree gum samples exhibited similar spots to the Chinese propolis samples, but the spots were less resolved, and the spots corresponding in Rf values to those in the chromatogram of S1 were missing. Using simple visual comparison, differences in the TLC chromatograms could be observed between the Chinese propolis samples and the poplar-tree gum samples. However, personal subjective judgments can vary and affect the impartiality of the discrimination. Therefore, it was necessary to create an objective computerised method for discrimination. The fluorescent image of the TLC plates exhibits a non-uniform background and should be corrected to obtain an accurate image analysis result (Zhang and Lin, 2006; Tang and Wu, 2008). For the images of the TLC plates captured under natural light, this issue was less pronounced. The three-dimensional view (Fig. 2) of the TLC chromatogram performed by person 1 revealed a relatively uniform background. Therefore, the image was analysed without background correction. Several pixel resolutions of the image were evaluated for image analysis. The similarities between densitograms T1 and T2, calculated according to Pearson’s similarity, were 0.9676, 0.9678, 0.9679, 0.9675 and 0.9721 for pixel resolutions of 500, 400, 300, 256 and 128, respectively. This result indicates that higher pixel resolution exerts only a minor effect on similarity. Because higher pixel resolutions lead to longer analysis times, an image height resolution of 256 pixels was chosen, which was sufficiently large and fit the binary number system of the computer.
Figure 2. The three-dimensional view of the thin-layer chromatogram performed by person 1.
Figure 1. Thin-layer chromatographic chromatograms of samples performed separately by three people: (A) person 1, (B) person 2 and (C) person 3. S1, quercetin; S2, kaempferol; S3, galangin; S4, chrysin; T1–T8, Chinese propolis samples; F1–F3, poplar-tree gum samples.
Phytochem. Anal. 2014
Copyright © 2014 John Wiley & Sons, Ltd.
wileyonlinelibrary.com/journal/pca
T.-x. Tang et al. Following the image analysis, densitograms were obtained for chemometric analysis. Figure 3 displays the densitograms of all 11 samples obtained from the TLC chromatogram performed by person 1. The reconstructed TLC chromatograms of the common pattern of densitograms for Chinese propolis and the sample densitograms are displayed in Fig. 4. The densitograms were aligned by offsetting. The variables of each densitogram were scaled to a range from 0 to 255 to ensure an accurate comparison between them and to facilitate the reconstruction of the TLC chromatogram. The reconstruction of the TLC chromatograms reduced the variance among the replicate measurements and enabled a clear visualisation of the chemical profile of each sample. For visual observation, colour is a critical factor in making discriminations. However, due to variations in visualisation and imaging, it was impossible to normalise the colour of the chromatograms using computer software. Instead, the colour of the TLC chromatograms was converted to greyscale to reduce variations and improve pattern recognition.
Chemometric analysis The chemometric methods were first screened according to two criteria: (i) the existence of a mature, accessible software tool for the method and (ii) whether additional data manipulation was
Figure 3. Densitograms of all 11 samples obtained from the TLC chromatogram performed by person 1. T1–T8, Chinese propolis samples; F1–F3, poplar-tree gum samples.
required. The chemometric methods were then further tested to determine whether all of the densitograms could be classified. Finally, SA, HC, KMC, NN and SVM were studied and compared. A common pattern of densitograms for Chinese propolis was generated for SA by calculating the mean of 24 densitograms from the Chinese propolis samples. The similarity between each densitogram and the common pattern was measured using cosine similarity and Pearson correlation. The highest similarity value of the densitograms of the poplar-tree gum samples was used as a threshold to discriminate the remaining densitograms. Using this criterion, a correct classification rate using cosine similarity of 97.0% was achieved, which was better than that achieved using Pearson correlation (Tables 2 and 3). The relative standard deviations (RSD) of the data for the cosine similarity method were also better (Tables 2 and 3). The misclassifications may have occurred because the densitograms were compared with the common pattern, and the similarities among the densitograms were not taken into account. The similar chemical profile of the counterfeit samples and the poor reproducibility of the TLC complicated the pattern recognition of the densitograms. Thus, it was challenging to define ‘true’ and ‘false’ conditions using SA. In HC analysis, the cosine similarity and the Pearson correlation were selected for ‘distance’ in SPSS. The defaults from the other options in SPSS were used. From Fig. 5, it is apparent that the distances between the densitograms (with the same analyst label) from different samples on the same plate were smaller than the distances between the densitograms (with the same sample code) for replicates of the same sample. This observation reveals the poor reproducibility of the TLC method, which complicated the pattern recognition of the densitograms. However, despite the poor TLC reproducibility, the samples were correctly classified into two main clusters comprising the Chinese propolis samples and the poplartree gum samples, regardless of whether the cosine similarity or the Pearson correlation was used to measure the distance (Fig. 5). k-Means clustering aims to classify objects into a known number of clusters, in which each object belongs to the cluster with the nearest mean (David, 2003). It is a more rapid clustering algorithm than HC. The KMC analysis using SPSS produced two clusters. Other options in SPSS were changed and tested. However, regardless of the options used, densitogram T3 (person 1) was always classified as a poplar-tree gum sample. The TLC chromatogram of sample T3 was lighter than those of the other samples (Fig. 1). However, the primary spots of sample T3 were in accordance with those of the Chinese propolis samples, and densitograms T3 (person 2) and T3 (person 3) were classified as Chinese propolis. Therefore, it was concluded that KMC did not provide adequate pattern recognition of the densitograms. This
Figure 4. Reconstructed thin-layer chromatograms of common pattern (generated by calculating the mean of 24 densitograms from the Chinese propolis samples) and the sample densitograms corresponding to Fig. 1. A, B and C correspond to TLC chromatogram performed by person 1, person 2 and person 3, respectively; T1–T8, Chinese propolis samples; F1–F3, poplar-tree gum samples.
wileyonlinelibrary.com/journal/pca
Copyright © 2014 John Wiley & Sons, Ltd.
Phytochem. Anal. 2014
Chemometric TLC Identification of Chinese Propolis Table 2. Cosine similarity between each densitogram and the common pattern TLC analyst
Person 1 Person 2 Person 3 RSD (%)
Sample codes T1
T2
T3
T4
T5
T6
T7
T8
F1
F2
F3
0.995 0.997 0.993 0.20
0.996 0.994 0.991 0.25
0.967 0.977 0.975 0.54
0.990 0.991 0.997 0.38
0.996 0.997 0.994 0.15
0.985 0.988 0.996 0.57
0.997 0.998 0.991 0.38
0.998 0.998 0.990 0.46
0.956 0.972 0.969 0.88
0.949 0.964 0.959 0.80
0.968 0.957 0.958 0.63
Bold face indicates the highest similarity value of the densitograms of the poplar-tree gum and the similarity values of the densitograms of Chinese propolis smaller than that value.
Table 3. Pearson correlation between each densitogram and the common pattern TLC analyst
Person 1 Person 2 Person 3 RSD (%)
Sample codes T1
T2
T3
T4
T5
T6
T7
T8
F1
F2
F3
0.942 0.965 0.972 1.6
0.971 0.972 0.927 2.7
0.894 0.868 0.809 5.1
0.870 0.900 0.962 5.2
0.953 0.964 0.983 1.6
0.791 0.838 0.953 9.7
0.965 0.975 0.957 0.9
0.971 0.983 0.937 2.5
0.438 0.612 0.661 20.5
0.685 0.818 0.825 10.2
0.721 0.756 0.764 3.1
Bold face indicates the highest similarity value of the densitograms of the poplar-tree gum and the similarity values of the densitograms of Chinese propolis smaller than that value.
Figure 5. Dendrogram of hierarchical clustering analysis. T1–T8, indicate Chinese propolis samples; F1–F3, indicate poplar-tree gum samples; person 1, person 2 and person 3, indicate the people who performed the TLC of the samples.
failure may have arisen because the KMC algorithm considers only the distance between the means and the objects, which ignores the weight and breadth of each cluster (David, 2003). Neural network is a supervised learning algorithm used for modelling complex relationships between inputs and outputs and has been used successfully for pattern recognition (David, 2003). In NN analysis using SPSS, both multilayer perception (MLP) and radial basis function (RBF) analyses were evaluated.
Phytochem. Anal. 2014
Initially, the densitograms were assigned randomly by SPSS to training, testing and holdout groups, with numbers of 22, 6 and 5, respectively. The holdout densitograms were not used to build the model in order to allow an accurate estimate of the predictive ability of the model. The classification results were selected for output, the predictive results were selected to be saved and the defaults of the other options were used. The MLP and RBF analyses were run five times. Densitogram
Copyright © 2014 John Wiley & Sons, Ltd.
wileyonlinelibrary.com/journal/pca
T.-x. Tang et al. T3 (person 1) was misclassified once in the five runs of the MLP and RBF analyses. Densitograms were subsequently manually assigned. It was determined that if all three of the T3 densitograms were assigned to the holdout group, the modelling was poor. Therefore, some options in SPSS were changed and tested. When the architecture of the MLP analysis was changed from an automatic selection to two hidden layers of a sigmoid activation function and an output layer of an identity activation function, the model produced with the T3 densitograms assigned to the holdout group achieved a correct classification rate of 100%. For the RBF analyses, no satisfactory results could be obtained by changing the options. The SVM is also a supervised learning algorithm and is useful for the pattern recognition of complex datasets (Vapnik, 1998). Initially, one densitogram was selected randomly from the three replicates for each sample and was assigned to testing, while the other two densitograms were assigned to training. The default options in the LIBSVM software were used. All five runs of the SVM analysis exhibited correct classification rates of 100%. Based on the experience from NN analysis, the SVM analysis was run again with all three densitograms from the T3 sample assigned to testing. Densitogram T3 (person 1) was again misclassified. The options were therefore changed and re-evaluated. When the kernel function was changed from RBF to sigmoid and the defaults for the other options were used, the model trained without the densitograms from the T3 sample achieved a correct classification rate of 100%. Although the similar chemical profile of the counterfeit samples and the poor reproducibility of the TLC made the pattern recognition of the densitograms challenging, HC, NN and SVM were able to achieve a correct classification rate of 100%, demonstrating that TLC combined with image analysis and chemometrics can be used for the identification of Chinese propolis.
A strategy for TLC identification of Chinese propolis using chemometric fingerprinting A strategy for TLC identification of Chinese propolis using chemometric fingerprinting is proposed, as depicted in Fig. 6. Densitograms are first obtained by image analysis. Densitograms of known samples are then used to produce a common pattern, construct a sample database and train the NN/SVM discrimination model. Unknown samples are recognised using (I) SA between the densitogram of the unknown sample and the common pattern, (II) HC analysis of the densitogram of the unknown sample and those of known samples in the database, and (III) NN/SVM classification of the densitogram of the unknown sample. As SA measures the similarity and supplies an objective, quantifiable criterion for quality control, it can be used to strictly control the consistency of products from the same origin. Hierarchical clustering is used as a non-supervised pattern recognition method. Neural network and SVM are used as supervised pattern recognition methods. The NN is based on empirical risk minimisation, while SVM is based on structural risk minimisation (Vapnik, 1998). These methods are complementary but not identical and are therefore both used in our classification strategy. Sample identification can be achieved using the results of all of the analyses, or those of a specific subset of the analyses. For the identification of Chinese propolis, SA is not recommended as an analytical tool, because the differences among Chinese propolis samples of varying origin are considerable and because the counterfeit samples exhibit similar chemical profiles, true Chinese propolis samples will occasionally be misclassified (false negative). In addition, counterfeit samples may occasionally be classified as propolis (false positive). Therefore, more flexible pattern recognition methods are necessary. Hierarchical clustering, NN and SVM analyses together provided more accurate sample identification. Agreement between the
Figure 6. Schematic diagram of the thin-layer chromatographic identification using chemometric fingerprinting. Densitograms are obtained by image analysis. Densitograms of known samples are used to produce a common pattern, construct a sample database and train the neural network/support vector machine discrimination model. Unknown samples are recognised using (I) similarity analysis, (II) hierarchical clustering, and (III) neural network/ support vector machine analyses.
wileyonlinelibrary.com/journal/pca
Copyright © 2014 John Wiley & Sons, Ltd.
Phytochem. Anal. 2014
Chemometric TLC Identification of Chinese Propolis results of these methods is necessary to ensure objective computerised discrimination. Our developed strategy can be implemented using all-in-one software to create a black-box system. The common pattern, sample database and pattern recognition models can be established by a regulatory body using known samples. Users would then be able to discriminate unknown samples using the software. Chemometric analysis of the image data allowed an objective interpretation of the TLC chromatograms from the samples. The strategy for TLC identification method using chemometric fingerprinting developed in this study provides a rapid, low-cost tool for the quality control of Chinese propolis and can be extended to the identification of other natural products with complex chemical matrices.
References Chang CC, Lin CJ. 2011. LIBSVM: a library for support vector machines. ACM Trans Intell Sys Technol 2: 1–27 [software available at http:// www.csie.ntu.edu.tw/~cjlin/libsvm]. Cheng H, Qin ZH, Guo XF, Hu XS, Wu JH. 2013. Geographical origin identification of propolis using GC–MS and electronic nose combined with principal components analysis. Food Res Int 51: 813–822. Committee of National Pharmacopeia. 2010. Pharmacopeia of People’s Republic of China (Part 1). People’s Medical Publishing House: Beijing; 336–337. David JCM. 2003. Information Theory, Inference and Learning Algorithms. Cambridge University Press: Cambridge; 286–289, 467–527. Gardana C, Scaglianti, M, Pietta P, Simonetti P. 2007. Analysis of the polyphenolic fraction of propolis from different sources by liquid chromatography–tandem mass spectrometry. J Pharm Biomed Anal 45: 390–399. Medana C, Carbone F, Aigotti R, Appendino G, Baiocchi C. 2008. Selective analysis of phenolic compounds in propolis by HPLC–MS/MS. Phytochem Anal 19: 32–39. Nagaraja N, Bankova VS, Afrouzan H. 2010. Fingerprinting of propolis by easy ambient sonic-spray ionization mass spectrometry. Talanta 81: 100–108.
Phytochem. Anal. 2014
Ogegbo OL, Eyob S, Parmar S, Wang Z, Bligha SWA. 2012. Metabolomics of four TCM herbal products: application of HPTLC analysis. Anal Methods 4: 2522–2527. Paviani LC, Dariva C, Marcucci MC, Cabral FA. 2010. Supercritical carbon dioxide selectivity to fractionate phenolic compounds from the dry ethanolic extract of propolis. J Food Process Eng 33: 15–27. Salatino A, Fernandes-Silva CC, Righi AA, Salatino MLF. 2011. Propolis research and the chemistry of plant products. Nat Prod Rep 28: 925–936. Sârbu C, Moţ AC. 2011. Ecosystem discrimination and fingerprinting of Romanian propolis by hierarchical fuzzy clustering and image analysis of TLC patterns. Talanta 85: 1112–1117. Sforcin JM, Bankova V. 2011. Propolis: Is there a potential for the development of new drugs? J Ethnopharmacol 133: 253–260. Stefano C, Francesco C. 2002. Propolis, an old remedy used in modern medicine. Fitoterapia 73(Suppl. 1): S1–S6. Tang TX, Wu H. 2008. An image analysis system for thin-layer chromatography quantification and its validation. J Chromatogr Sci 46: 560–564. Tian RT, Xie PS, Liu HP. 2009. Evaluation of traditional Chinese herbal medicine: Chaihu (Bupleuri radix) by both high-performance liquid chromatographic and high-performance thin-layer chromatographic fingerprint and chemometric analysis. J Chromatogr A 1216: 2150–2155. Vapnik VN. 1998. Statistical Learning Theory. Wiley: New York; 219–221, 493–503. Wu YW, Sun SQ, Zhao J, Li Y, Zhou Q. 2008. Rapid discrimination of extracts of Chinese propolis and poplar buds by FT-IR and 2D IR correlation spectroscopy. J Mol Struct 883–884: 48–54. Zarzycki PK, Zarzycka MB, Clifton VL, Adamski J, Głód BK. 2011. Lowparachor solvents extraction and thermostated micro-thin-layer chromatography separation for fast screening and classification of spirulina from pharmaceutical formulations and food samples. J Chromatogr A 1218: 5694–5704. Zhang L, Lin X. 2006. Quantitative evaluation of thin-layer chromatography with image background estimation based on charge coupled device imaging. J Chromatogr A 1109: 273–278. Zhang CP, Zheng HQ, Liu G, Hu FL. 2011. Development and validation of HPLC method for determination of salicin in poplar buds: Application for screening of counterfeit propolis. Food Chem 127: 345–350. Zhou JH, Li Y, Zhao J, Xue XF, Wu LM, Chen F. 2008. Geographical traceability of propolis by high-performance liquid-chromatography fingerprints. Food Chem 108: 749–759.
Copyright © 2014 John Wiley & Sons, Ltd.
wileyonlinelibrary.com/journal/pca
Revised: 28 November 2013,
Accepted: 30 November 2013
Published online in Wiley Online Library
(wileyonlinelibrary.com) DOI 10.1002/pca.2502
Thin-layer Chromatographic Identification of Chinese Propolis Using Chemometric Fingerprinting Tie-xin Tang,a Wei-yan Guo,a Ye Xu,a Si-ming Zhang,a Xin-jun Xu,a,b Dong-mei Wang,a Zhi-min Zhao,a,b Long-ping Zhua,b and De-po Yanga,b* ABSTRACT: Introduction – Poplar tree gum has a similar chemical composition and appearance to Chinese propolis (bee glue) and has been widely used as a counterfeit propolis because Chinese propolis is typically the poplar-type propolis, the chemical composition of which is determined mainly by the resin of poplar trees. The discrimination of Chinese propolis from poplar tree gum is a challenging task. Objective – To develop a rapid thin-layer chromatographic (TLC) identification method using chemometric fingerprinting to discriminate Chinese propolis from poplar tree gum. Methods – A new TLC method using a combination of ammonia and hydrogen peroxide vapours as the visualisation reagent was developed to characterise the chemical profile of Chinese propolis. Three separate people performed TLC on eight Chinese propolis samples and three poplar tree gum samples of varying origins. Five chemometric methods, including similarity analysis, hierarchical clustering, k-means clustering, neural network and support vector machine, were compared for use in classifying the samples based on their densitograms obtained from the TLC chromatograms via image analysis. Results – Hierarchical clustering, neural network and support vector machine analyses achieved a correct classification rate of 100% in classifying the samples. A strategy for TLC identification of Chinese propolis using chemometric fingerprinting was proposed and it provided accurate sample classification. Conclusion – The study has shown that the TLC identification method using chemometric fingerprinting is a rapid, low-cost method for the discrimination of Chinese propolis from poplar tree gum and may be used for the quality control of Chinese propolis. Copyright © 2014 John Wiley & Sons, Ltd. Keywords: chemometrics; fingerprinting; identification; image analysis; pattern recognition; thin-layer chromatography; Chinese propolis
Introduction Propolis is produced by worker bees, which collect resin from plants, mix it with their salivary secretions, and finally add beeswax to produce the propolis (bee glue). Propolis has been marketed and used extensively in health foods, cosmetics and traditional medicines for many years. It exhibits numerous beneficial bioactivities, including anti-inflammatory, anti-oxidant, anti-bacterial, anti-viral, anti-fungal, immunomodulatory and anti-tumour effects (Stefano and Francesco, 2002; Salatino et al., 2011; Sforcin and Bankova, 2011). In Chinese pharmacopeia, propolis is listed as a drug for treating hyperlipidaemia, diabetes, etc., in the volume on traditional Chinese medicine (Committee of National Pharmacopeia, 2010). Plant resins are the primary material used by bees to produce propolis and they determine its chemical composition. The majority of Chinese propolis originates from the Populus species and its hybrids. Chinese propolis is typically classified as poplartype (Wu et al., 2008). Its major bioactive components are polyphenols such as flavonoids and caffeic-acid derivatives (Gardana et al., 2007; Medana et al., 2008; Salatino et al., 2011). Poplar tree gum, the extract of poplar buds, has a similar chemical composition and appearance to Chinese propolis and has been widely used as a counterfeit (Wu et al., 2008; Zhang et al., 2011). China is the world’s second largest producer of propolis (Paviani et al., 2010), and it
Phytochem. Anal. 2014
is extremely important to discriminate Chinese propolis from poplar tree gum. Although numerous studies have addressed the discrimination and fingerprinting of propolis samples (Gardana et al., 2007; Medana et al., 2008; Zhou et al., 2008; Nagaraja et al., 2010; Sârbu and Moţ, 2011; Cheng et al., 2013), few have specifically tackled the discrimination of Chinese propolis from poplar tree gum. A method employing Fourier transform infrared spectroscopy combined with two-dimensional infrared correlation analysis was developed to distinguish the extracts of Chinese propolis from the extracts of poplar buds (Wu et al., 2008). In another approach, HPLC of salicin was used to screen for counterfeit Chinese propolis (Zhang et al., 2011). Unfortunately, these methods require the use of expensive instrumentation.
* Correspondence to: De-po Yang, School of Pharmaceutical Sciences, Sun Yat-sen University, Guangzhou, China 510006. E-mail: [email protected] a
School of Pharmaceutical Sciences, Sun Yat-sen University, Guangzhou, China, 510006
b
Guangdong Technology Research Center for Advanced Chinese Medicine, Guangzhou, China, 510006
Copyright © 2014 John Wiley & Sons, Ltd.
T.-x. Tang et al. In contrast, TLC is simple, flexible, accessible and affordable. It is the most commonly used identification method for herbal medicines in the pharmacopeias of countries all over the world. Thin-layer chromatographic fingerprinting is most often conducted in a subjective manner using manually determined peak differences, and the application of chemometrics to TLC classification is still under development. However, TLC chromatograms can be captured using a digital camera, and image and chemometric analyses can be performed on a personal computer. The combination of TLC with image analysis and chemometrics can provide a fascinating, rapid and low-cost analytical tool for the identification of natural products. Several studies have reported the application of chemometrics to TLC using image analysis for data acquisition. Principal component analysis (PCA), partial-least-squares discriminant analysis (PLS-DA) and orthogonal PLS-DA were used for a metabolomics study of the HPTLC fingerprints of four medicinal herbs (Ogegbo et al., 2012). Thin-layer chromatography was combined with image analysis and chemometric analyses, including similarity analysis (SA), artificial neural network (ANN, otherwise known as neural network, NN), and k-nearest neighbours (k-NNs), to classify samples of the genus Bupleuri radix (Tian et al., 2009). Principal components analysis and hierarchical clustering (HC) analysis were used with thermostated microthin-layer chromatography for the rapid screening and classification of spirulina from pharmaceutical formulations and food samples (Zarzycki et al., 2011). Thin-layer chromatography was combined with image analysis and hierarchical fuzzy clustering for the successful ecosystem discrimination and fingerprinting of Romanian propolis (Sârbu and Moţ, 2011). In the above cases, although the sample separation on the TLC plate was not good, the chemical profile of the samples resulted in the formation of unique features on the TLC chromatograms. Using chemometric methods, the samples could be classified correctly based on the TLC results, suggesting that TLC combined with image analysis and chemometrics can be used for the identification of Chinese propolis. In this study, a new TLC method for the identification of Chinese propolis was developed. The chemometric methods SA, HC, k-means clustering (KMC), NN and support vector machine (SVM) were compared for use in classifying the samples based on their TLC densitograms. Furthermore, a strategy to discriminate between Chinese propolis and poplar tree gum samples is proposed.
Experimental Materials and reagents Chinese propolis samples and poplar-tree gum samples were collected from beekeepers or supplied by commercial companies in China. The details of the samples studied are provided in Table 1. Reference standards for quercetin (S1), kaempferol (S2), galangin (S3) and chrysin (S4) were purchased from the National Institute for the Control of Pharmaceutical and Biological Products (Beijing, China). Methanol, toluene, ethyl acetate, formic acid, ammonia aqueous solution (25%) and hydrogen peroxide aqueous solution (30%), from the Guangzhou Chemical Reagent Factory (Guangzhou, China), were of analytical grade. Silica gel 60 pre-coated TLC plates (100 × 200 mm) were purchased from Qingdao Haiyang Chemical Co., Ltd (Qingdao, China). The samples and reference standards were weighed using an AG285 analytical balance (Mettler Toledo Instruments AG, Zurich, Switzerland). Ultrasonic extraction was accomplished using an SB25-12D ultrasonic cleaner (Ningbo Scientz Bio-tech Co., Ltd, Ningbo, China). The solutions were applied to the TLC plate using a microlitre syringe (Shanghai GaoKe Industrial and Trading Co., Ltd, Shanghai, China). The plates were developed in a 200 × 200 mm (length × height) twin-trough glass chamber (Shanghai Xinyi Instrument Co., Ltd, Shanghai, China). The TLC images were captured using an A700 digital camera (Cannon Inc., Tokyo, Japan).
Extraction and thin-layer chromatography Accurately weighed 0.5 g samples were extracted with 20 mL of methanol for 20 min by ultrasonic extraction. The extract solution was filtered and used as the sample solution. Accurately weighed quercetin, kaempferol, galangin and chrysin were dissolved in appropriate volumes of methanol to produce solutions of 0.254, 0.258, 0.255 and 0.251 mg/mL, respectively. The sample solutions and standard solutions were applied in spots on a TLC plate. A mixture of toluene:ethyl acetate:formic acid (10:4:0.5) was used as the mobile phase. A twin-trough chamber was presaturated with the mobile phase for 15 min, and the plate was developed in the chamber until the solvent front had migrated approximately 7 cm from the origin. The plate was then removed from the chamber, dried in an oven at 105 °C for 10 min, and placed on a rack in another twin-trough chamber. An ammonia aqueous solution of 10 mL was added to one trough of the second chamber, and a hydrogen peroxide aqueous solution of 10 mL was added to the other trough. The second chamber was covered with a lid, and the plate was treated with the ammonia and hydrogen peroxide vapours until the colour of the spots on the plate had stabilised. Finally, the plate was removed and photographed under natural light. Three separate people performed the TLC analysis in order to obtain more data to train and test the discrimination models in chemometric fingerprinting.
Table 1. Origins of the Chinese propolis and poplar-tree gum samples Samples Chinese propolis Chinese propolis Chinese propolis Chinese propolis Chinese propolis Chinese propolis Chinese propolis Chinese propolis Poplar tree gum Poplar tree gum Poplar tree gum
Sample codes
Suppliers
T1 T2 T3 T4 T5 T6 T7 T8 F1 F2 F3
Xiangyun bee yard, Xuancheng, Anhui Zhuxuefeng bee yard, Qingdao, Shandong Lengjiao bee-related products store, Meishan, Sichuan Purui bee-related products company, Xi’an, Shangxi Yifengtang bee-related products store, Xi’an, Shangxi Luohandong bee yard, Dongda county, Xi’an, Shangxi Dahua bee-related products company, Zhengzhou, Henan (collected from Shandong) Dahua bee-related products company, Zhengzhou, Henan (collected from Anhui) Xiemei bee-related products company, Shanghai Xilian bee-related products company, Lanzhou, Gansu Dahua bee-related products company, Zhengzhou, Henan
wileyonlinelibrary.com/journal/pca
Copyright © 2014 John Wiley & Sons, Ltd.
Phytochem. Anal. 2014
Chemometric TLC Identification of Chinese Propolis Image analysis Images of the TLC chromatograms, saved in JPEG format, were analysed using the self-programmed software developed by Delphi Xe2 IDE (Embarcadero, San Francisco, CA, USA). First, the image was pre-treated by manually rotating to horizontally align the initial sample spots. The image was then manually cropped from the initial sample spots to the solvent front to eliminate the irrelevant regions. Subsequently, the image was automatically converted to greyscale and scaled to a height of 256 pixels. Second, the image was virtually scanned and automatically segmented into lanes corresponding to the sample migration paths. By summing the grey values of the pixels in each row at the same migration distance in the lane, a one-dimensional array was automatically obtained. A densitogram was then produced by plotting the line graph on an axis with the data of the one-dimensional array. Finally, the densitograms were manually offset by fixing the Rf of common compounds to the same value (Sârbu and Moţ, 2011), and the numeric variables of each densitogram were automatically scaled in the range from 0 to 255. Thin-layer chromatographic greyscale chromatograms reconstructed from the densitograms, were created by applying the densitogram data as the greyscale value to the pixels of a line, and repeating the line 28 times.
Chemometric analysis One densitogram (256 variables) was used as a dataset for chemometric analysis. A total of 33 densitograms (11 samples × 3) constituted the whole datasets. Similarity, HC, KMC and NN analyses were performed by SPSS (Version 21, IBM Corporation, Armonk, New York, USA). Support vector machine analysis was performed by LIBSVM (Version 3.16) (Chang and Lin, 2011). The options were optimised to obtain the best results.
Results and discussion Thin-layer chromatography of the samples and its image analysis A toluene:ethyl acetate:formic acid (30:12:5, v/v/v) mixture was used as the mobile phase for the TLC of Romanian propolis samples according to the previous method (Sârbu and Moţ, 2011). However, this mixture was not suitable for the TLC of Chinese propolis using silica gel 60 pre-coated plates. The solvent ratio was therefore adjusted, and a toluene:ethyl acetate:formic acid ratio of 10:4:0.5 was found to provide improved separation. In the previous method (Sârbu and Moţ, 2011) the TLC of Romania propolis samples was visualised by spraying a methanolic solution of 0.2% diphenylboryloxyethylamine and 4% polyethyleneglycol on the plate and capturing a fluorescent image under UV 366 nm light using a UV viewer cabinet. To circumvent the need to use fluorescent imaging equipment, a new visualisation method was developed using ammonia and
hydrogen peroxide vapours. The hydrogen peroxide was used as the oxidant, and the ammonia vapour ionised the polyphenols, increasing their susceptibility to oxidation. The TLC chromatograms of the samples, performed separately by three individuals, are displayed in Fig. 1. The polyphenols were oxidised by the ammonia and hydrogen peroxide vapours and appear different in colour. There were five main spots in the chromatograms of the Chinese propolis samples. Four of the spots corresponded in Rf values to those in the chromatograms of S1, S2, S3 and S4. The chromatograms of the three poplar-tree gum samples exhibited similar spots to the Chinese propolis samples, but the spots were less resolved, and the spots corresponding in Rf values to those in the chromatogram of S1 were missing. Using simple visual comparison, differences in the TLC chromatograms could be observed between the Chinese propolis samples and the poplar-tree gum samples. However, personal subjective judgments can vary and affect the impartiality of the discrimination. Therefore, it was necessary to create an objective computerised method for discrimination. The fluorescent image of the TLC plates exhibits a non-uniform background and should be corrected to obtain an accurate image analysis result (Zhang and Lin, 2006; Tang and Wu, 2008). For the images of the TLC plates captured under natural light, this issue was less pronounced. The three-dimensional view (Fig. 2) of the TLC chromatogram performed by person 1 revealed a relatively uniform background. Therefore, the image was analysed without background correction. Several pixel resolutions of the image were evaluated for image analysis. The similarities between densitograms T1 and T2, calculated according to Pearson’s similarity, were 0.9676, 0.9678, 0.9679, 0.9675 and 0.9721 for pixel resolutions of 500, 400, 300, 256 and 128, respectively. This result indicates that higher pixel resolution exerts only a minor effect on similarity. Because higher pixel resolutions lead to longer analysis times, an image height resolution of 256 pixels was chosen, which was sufficiently large and fit the binary number system of the computer.
Figure 2. The three-dimensional view of the thin-layer chromatogram performed by person 1.
Figure 1. Thin-layer chromatographic chromatograms of samples performed separately by three people: (A) person 1, (B) person 2 and (C) person 3. S1, quercetin; S2, kaempferol; S3, galangin; S4, chrysin; T1–T8, Chinese propolis samples; F1–F3, poplar-tree gum samples.
Phytochem. Anal. 2014
Copyright © 2014 John Wiley & Sons, Ltd.
wileyonlinelibrary.com/journal/pca
T.-x. Tang et al. Following the image analysis, densitograms were obtained for chemometric analysis. Figure 3 displays the densitograms of all 11 samples obtained from the TLC chromatogram performed by person 1. The reconstructed TLC chromatograms of the common pattern of densitograms for Chinese propolis and the sample densitograms are displayed in Fig. 4. The densitograms were aligned by offsetting. The variables of each densitogram were scaled to a range from 0 to 255 to ensure an accurate comparison between them and to facilitate the reconstruction of the TLC chromatogram. The reconstruction of the TLC chromatograms reduced the variance among the replicate measurements and enabled a clear visualisation of the chemical profile of each sample. For visual observation, colour is a critical factor in making discriminations. However, due to variations in visualisation and imaging, it was impossible to normalise the colour of the chromatograms using computer software. Instead, the colour of the TLC chromatograms was converted to greyscale to reduce variations and improve pattern recognition.
Chemometric analysis The chemometric methods were first screened according to two criteria: (i) the existence of a mature, accessible software tool for the method and (ii) whether additional data manipulation was
Figure 3. Densitograms of all 11 samples obtained from the TLC chromatogram performed by person 1. T1–T8, Chinese propolis samples; F1–F3, poplar-tree gum samples.
required. The chemometric methods were then further tested to determine whether all of the densitograms could be classified. Finally, SA, HC, KMC, NN and SVM were studied and compared. A common pattern of densitograms for Chinese propolis was generated for SA by calculating the mean of 24 densitograms from the Chinese propolis samples. The similarity between each densitogram and the common pattern was measured using cosine similarity and Pearson correlation. The highest similarity value of the densitograms of the poplar-tree gum samples was used as a threshold to discriminate the remaining densitograms. Using this criterion, a correct classification rate using cosine similarity of 97.0% was achieved, which was better than that achieved using Pearson correlation (Tables 2 and 3). The relative standard deviations (RSD) of the data for the cosine similarity method were also better (Tables 2 and 3). The misclassifications may have occurred because the densitograms were compared with the common pattern, and the similarities among the densitograms were not taken into account. The similar chemical profile of the counterfeit samples and the poor reproducibility of the TLC complicated the pattern recognition of the densitograms. Thus, it was challenging to define ‘true’ and ‘false’ conditions using SA. In HC analysis, the cosine similarity and the Pearson correlation were selected for ‘distance’ in SPSS. The defaults from the other options in SPSS were used. From Fig. 5, it is apparent that the distances between the densitograms (with the same analyst label) from different samples on the same plate were smaller than the distances between the densitograms (with the same sample code) for replicates of the same sample. This observation reveals the poor reproducibility of the TLC method, which complicated the pattern recognition of the densitograms. However, despite the poor TLC reproducibility, the samples were correctly classified into two main clusters comprising the Chinese propolis samples and the poplartree gum samples, regardless of whether the cosine similarity or the Pearson correlation was used to measure the distance (Fig. 5). k-Means clustering aims to classify objects into a known number of clusters, in which each object belongs to the cluster with the nearest mean (David, 2003). It is a more rapid clustering algorithm than HC. The KMC analysis using SPSS produced two clusters. Other options in SPSS were changed and tested. However, regardless of the options used, densitogram T3 (person 1) was always classified as a poplar-tree gum sample. The TLC chromatogram of sample T3 was lighter than those of the other samples (Fig. 1). However, the primary spots of sample T3 were in accordance with those of the Chinese propolis samples, and densitograms T3 (person 2) and T3 (person 3) were classified as Chinese propolis. Therefore, it was concluded that KMC did not provide adequate pattern recognition of the densitograms. This
Figure 4. Reconstructed thin-layer chromatograms of common pattern (generated by calculating the mean of 24 densitograms from the Chinese propolis samples) and the sample densitograms corresponding to Fig. 1. A, B and C correspond to TLC chromatogram performed by person 1, person 2 and person 3, respectively; T1–T8, Chinese propolis samples; F1–F3, poplar-tree gum samples.
wileyonlinelibrary.com/journal/pca
Copyright © 2014 John Wiley & Sons, Ltd.
Phytochem. Anal. 2014
Chemometric TLC Identification of Chinese Propolis Table 2. Cosine similarity between each densitogram and the common pattern TLC analyst
Person 1 Person 2 Person 3 RSD (%)
Sample codes T1
T2
T3
T4
T5
T6
T7
T8
F1
F2
F3
0.995 0.997 0.993 0.20
0.996 0.994 0.991 0.25
0.967 0.977 0.975 0.54
0.990 0.991 0.997 0.38
0.996 0.997 0.994 0.15
0.985 0.988 0.996 0.57
0.997 0.998 0.991 0.38
0.998 0.998 0.990 0.46
0.956 0.972 0.969 0.88
0.949 0.964 0.959 0.80
0.968 0.957 0.958 0.63
Bold face indicates the highest similarity value of the densitograms of the poplar-tree gum and the similarity values of the densitograms of Chinese propolis smaller than that value.
Table 3. Pearson correlation between each densitogram and the common pattern TLC analyst
Person 1 Person 2 Person 3 RSD (%)
Sample codes T1
T2
T3
T4
T5
T6
T7
T8
F1
F2
F3
0.942 0.965 0.972 1.6
0.971 0.972 0.927 2.7
0.894 0.868 0.809 5.1
0.870 0.900 0.962 5.2
0.953 0.964 0.983 1.6
0.791 0.838 0.953 9.7
0.965 0.975 0.957 0.9
0.971 0.983 0.937 2.5
0.438 0.612 0.661 20.5
0.685 0.818 0.825 10.2
0.721 0.756 0.764 3.1
Bold face indicates the highest similarity value of the densitograms of the poplar-tree gum and the similarity values of the densitograms of Chinese propolis smaller than that value.
Figure 5. Dendrogram of hierarchical clustering analysis. T1–T8, indicate Chinese propolis samples; F1–F3, indicate poplar-tree gum samples; person 1, person 2 and person 3, indicate the people who performed the TLC of the samples.
failure may have arisen because the KMC algorithm considers only the distance between the means and the objects, which ignores the weight and breadth of each cluster (David, 2003). Neural network is a supervised learning algorithm used for modelling complex relationships between inputs and outputs and has been used successfully for pattern recognition (David, 2003). In NN analysis using SPSS, both multilayer perception (MLP) and radial basis function (RBF) analyses were evaluated.
Phytochem. Anal. 2014
Initially, the densitograms were assigned randomly by SPSS to training, testing and holdout groups, with numbers of 22, 6 and 5, respectively. The holdout densitograms were not used to build the model in order to allow an accurate estimate of the predictive ability of the model. The classification results were selected for output, the predictive results were selected to be saved and the defaults of the other options were used. The MLP and RBF analyses were run five times. Densitogram
Copyright © 2014 John Wiley & Sons, Ltd.
wileyonlinelibrary.com/journal/pca
T.-x. Tang et al. T3 (person 1) was misclassified once in the five runs of the MLP and RBF analyses. Densitograms were subsequently manually assigned. It was determined that if all three of the T3 densitograms were assigned to the holdout group, the modelling was poor. Therefore, some options in SPSS were changed and tested. When the architecture of the MLP analysis was changed from an automatic selection to two hidden layers of a sigmoid activation function and an output layer of an identity activation function, the model produced with the T3 densitograms assigned to the holdout group achieved a correct classification rate of 100%. For the RBF analyses, no satisfactory results could be obtained by changing the options. The SVM is also a supervised learning algorithm and is useful for the pattern recognition of complex datasets (Vapnik, 1998). Initially, one densitogram was selected randomly from the three replicates for each sample and was assigned to testing, while the other two densitograms were assigned to training. The default options in the LIBSVM software were used. All five runs of the SVM analysis exhibited correct classification rates of 100%. Based on the experience from NN analysis, the SVM analysis was run again with all three densitograms from the T3 sample assigned to testing. Densitogram T3 (person 1) was again misclassified. The options were therefore changed and re-evaluated. When the kernel function was changed from RBF to sigmoid and the defaults for the other options were used, the model trained without the densitograms from the T3 sample achieved a correct classification rate of 100%. Although the similar chemical profile of the counterfeit samples and the poor reproducibility of the TLC made the pattern recognition of the densitograms challenging, HC, NN and SVM were able to achieve a correct classification rate of 100%, demonstrating that TLC combined with image analysis and chemometrics can be used for the identification of Chinese propolis.
A strategy for TLC identification of Chinese propolis using chemometric fingerprinting A strategy for TLC identification of Chinese propolis using chemometric fingerprinting is proposed, as depicted in Fig. 6. Densitograms are first obtained by image analysis. Densitograms of known samples are then used to produce a common pattern, construct a sample database and train the NN/SVM discrimination model. Unknown samples are recognised using (I) SA between the densitogram of the unknown sample and the common pattern, (II) HC analysis of the densitogram of the unknown sample and those of known samples in the database, and (III) NN/SVM classification of the densitogram of the unknown sample. As SA measures the similarity and supplies an objective, quantifiable criterion for quality control, it can be used to strictly control the consistency of products from the same origin. Hierarchical clustering is used as a non-supervised pattern recognition method. Neural network and SVM are used as supervised pattern recognition methods. The NN is based on empirical risk minimisation, while SVM is based on structural risk minimisation (Vapnik, 1998). These methods are complementary but not identical and are therefore both used in our classification strategy. Sample identification can be achieved using the results of all of the analyses, or those of a specific subset of the analyses. For the identification of Chinese propolis, SA is not recommended as an analytical tool, because the differences among Chinese propolis samples of varying origin are considerable and because the counterfeit samples exhibit similar chemical profiles, true Chinese propolis samples will occasionally be misclassified (false negative). In addition, counterfeit samples may occasionally be classified as propolis (false positive). Therefore, more flexible pattern recognition methods are necessary. Hierarchical clustering, NN and SVM analyses together provided more accurate sample identification. Agreement between the
Figure 6. Schematic diagram of the thin-layer chromatographic identification using chemometric fingerprinting. Densitograms are obtained by image analysis. Densitograms of known samples are used to produce a common pattern, construct a sample database and train the neural network/support vector machine discrimination model. Unknown samples are recognised using (I) similarity analysis, (II) hierarchical clustering, and (III) neural network/ support vector machine analyses.
wileyonlinelibrary.com/journal/pca
Copyright © 2014 John Wiley & Sons, Ltd.
Phytochem. Anal. 2014
Chemometric TLC Identification of Chinese Propolis results of these methods is necessary to ensure objective computerised discrimination. Our developed strategy can be implemented using all-in-one software to create a black-box system. The common pattern, sample database and pattern recognition models can be established by a regulatory body using known samples. Users would then be able to discriminate unknown samples using the software. Chemometric analysis of the image data allowed an objective interpretation of the TLC chromatograms from the samples. The strategy for TLC identification method using chemometric fingerprinting developed in this study provides a rapid, low-cost tool for the quality control of Chinese propolis and can be extended to the identification of other natural products with complex chemical matrices.
References Chang CC, Lin CJ. 2011. LIBSVM: a library for support vector machines. ACM Trans Intell Sys Technol 2: 1–27 [software available at http:// www.csie.ntu.edu.tw/~cjlin/libsvm]. Cheng H, Qin ZH, Guo XF, Hu XS, Wu JH. 2013. Geographical origin identification of propolis using GC–MS and electronic nose combined with principal components analysis. Food Res Int 51: 813–822. Committee of National Pharmacopeia. 2010. Pharmacopeia of People’s Republic of China (Part 1). People’s Medical Publishing House: Beijing; 336–337. David JCM. 2003. Information Theory, Inference and Learning Algorithms. Cambridge University Press: Cambridge; 286–289, 467–527. Gardana C, Scaglianti, M, Pietta P, Simonetti P. 2007. Analysis of the polyphenolic fraction of propolis from different sources by liquid chromatography–tandem mass spectrometry. J Pharm Biomed Anal 45: 390–399. Medana C, Carbone F, Aigotti R, Appendino G, Baiocchi C. 2008. Selective analysis of phenolic compounds in propolis by HPLC–MS/MS. Phytochem Anal 19: 32–39. Nagaraja N, Bankova VS, Afrouzan H. 2010. Fingerprinting of propolis by easy ambient sonic-spray ionization mass spectrometry. Talanta 81: 100–108.
Phytochem. Anal. 2014
Ogegbo OL, Eyob S, Parmar S, Wang Z, Bligha SWA. 2012. Metabolomics of four TCM herbal products: application of HPTLC analysis. Anal Methods 4: 2522–2527. Paviani LC, Dariva C, Marcucci MC, Cabral FA. 2010. Supercritical carbon dioxide selectivity to fractionate phenolic compounds from the dry ethanolic extract of propolis. J Food Process Eng 33: 15–27. Salatino A, Fernandes-Silva CC, Righi AA, Salatino MLF. 2011. Propolis research and the chemistry of plant products. Nat Prod Rep 28: 925–936. Sârbu C, Moţ AC. 2011. Ecosystem discrimination and fingerprinting of Romanian propolis by hierarchical fuzzy clustering and image analysis of TLC patterns. Talanta 85: 1112–1117. Sforcin JM, Bankova V. 2011. Propolis: Is there a potential for the development of new drugs? J Ethnopharmacol 133: 253–260. Stefano C, Francesco C. 2002. Propolis, an old remedy used in modern medicine. Fitoterapia 73(Suppl. 1): S1–S6. Tang TX, Wu H. 2008. An image analysis system for thin-layer chromatography quantification and its validation. J Chromatogr Sci 46: 560–564. Tian RT, Xie PS, Liu HP. 2009. Evaluation of traditional Chinese herbal medicine: Chaihu (Bupleuri radix) by both high-performance liquid chromatographic and high-performance thin-layer chromatographic fingerprint and chemometric analysis. J Chromatogr A 1216: 2150–2155. Vapnik VN. 1998. Statistical Learning Theory. Wiley: New York; 219–221, 493–503. Wu YW, Sun SQ, Zhao J, Li Y, Zhou Q. 2008. Rapid discrimination of extracts of Chinese propolis and poplar buds by FT-IR and 2D IR correlation spectroscopy. J Mol Struct 883–884: 48–54. Zarzycki PK, Zarzycka MB, Clifton VL, Adamski J, Głód BK. 2011. Lowparachor solvents extraction and thermostated micro-thin-layer chromatography separation for fast screening and classification of spirulina from pharmaceutical formulations and food samples. J Chromatogr A 1218: 5694–5704. Zhang L, Lin X. 2006. Quantitative evaluation of thin-layer chromatography with image background estimation based on charge coupled device imaging. J Chromatogr A 1109: 273–278. Zhang CP, Zheng HQ, Liu G, Hu FL. 2011. Development and validation of HPLC method for determination of salicin in poplar buds: Application for screening of counterfeit propolis. Food Chem 127: 345–350. Zhou JH, Li Y, Zhao J, Xue XF, Wu LM, Chen F. 2008. Geographical traceability of propolis by high-performance liquid-chromatography fingerprints. Food Chem 108: 749–759.
Copyright © 2014 John Wiley & Sons, Ltd.
wileyonlinelibrary.com/journal/pca
