www.jss-journal.com
Page 1
Journal of Separation Science
Rapid discrimination of slimming capsules based on illegal additives by electronic nose and flash gas chromatography
Zhenzhen Xia, Wensheng Cai, Xueguang Shao Collaborative Innovation Center of Chemical Science and Engineering (Tianjin), State Key Laboratory of Medicinal Chemical Biology, and Research Center for Analytical Sciences, College of Chemistry, Nankai University, Tianjin, 300071, China
Keywords: Counterfeit medicines / Data fusion / Electronic nose / Flash gas chromatography / Linear discriminant analysis
Correspondence: Pro. Xueguang Shao, College of Chemistry, Nankai University, Tianjin, 300071, P. R. China Fax: +86-22-23502458 E-mail: [email protected]
Abbreviations: E-nose, electronic nose; GC, gas chromatography; PCA, principal component analysis; LDA, linear discriminant analysis; Received: 25-Aug-2014; Revised: 07-Nov-2014; Accepted: 19-Nov-2014 This article has been accepted for publication and undergone full peer review but has not been through the copyediting, typesetting, pagination and proofreading process, which may lead to differences between this version and the Version of Record. Please cite this article as doi: 10.1002/jssc.201400941. This article is protected by copyright. All rights reserved.
www.jss-journal.com
Page 2
Journal of Separation Science
Abstract The discrimination of counterfeit and/or illegally manufactured medicines is an important task in the pharmaceutical industry for pharmaceutical safety. In this study, 22 slimming capsule samples with illegally added sibutramine and phenolphthalein were analyzed by electronic nose and flash gas chromatography. To reveal the difference among the different classes of samples, the principal component analysis and linear discriminant analysis were employed to analyze the data acquired from electronic nose and flash gas chromatography, respectively. The samples without illegal additives can be discriminated from the ones with illegal additives by using electronic nose or flash gas chromatography data individually. To improve the performance of classification, data fusion strategy was applied to integrate the data from electronic nose and flash gas chromatography data into a single model. The results show that the samples with phenolphthalein, sibutramine and both can be classified well by using fused data.
1. Introduction Excess weight and obesity have been recognized as serious worldwide problems leading to the search for a number of weight-loss medicines, especially alternative treatments based on natural slimming formulations [1]. Although natural slimming formulations present some side effects, it is a fact that the potency of these products is lower than synthetic slimming agents [2,3]. As a result, some manufacturers committed illegal adulterations of synthetic anorexics and cathartics to improve the efficacy of their products. Phenolphthalein and sibutramine are often added illegally into natural slimming products [4]. Phenolphthalein is This article is protected by copyright. All rights reserved.
www.jss-journal.com
Page 3
Journal of Separation Science
an acid–base indicator and laxative, which is prescribed for short-term treatment of constipation in humans, implying its possible weight-reducing effect. However, it can cause more severe constipation, electrolyte imbalance, cancer, cardiac arrhythmia and even death [5]. Sibutramine is a serotonin noradrenalin reuptake inhibitor, which promotes and maintains weight loss in obese people. However, it simultaneously overexcites the central nervous system and several side effects like nervousness, xerostomia, headache, numbness and paraesthesia have been reported [6]. Counterfeit medicines with illegal additives pose a huge threat to public health worldwide. Hence, the discrimination of counterfeit medicines has been an important task in pharmaceutical industry [7,8]. Odors of natural slimming medicine could vary between genuine and counterfeit medicines [9]. Therefore, to detect the subtle differences of odors is a feasible way to distinguish natural slimming medicine. The electronic nose (E-nose) [10], engineered to mimic the mammalian olfactory system, is designed to acquire objective data and perform odor evaluation on a continuous basis with a minimal cost. It identifies the mixture of volatile constituents as a whole without having to detect the individual chemical constituents within the mixture. Its array of cross-reactive sensors can interact with a broad range of volatile compounds, and the multivariate responses can be used as an “electronic fingerprint” to identify the odors through pattern recognition. Its advantages also include no requirement of sample pretreatment and rapid and sensitive analysis [11–13]. Electronic noses have been applied successfully in natural medicine analysis [14]. Flash GC also offers an effective and rapid tool for the volatiles characterization and detection in the different area [15–19], like pharmaceutical analysis, food analysis, environmental analysis, etc. Besides, chemometrics This article is protected by copyright. All rights reserved.
www.jss-journal.com
Page 4
Journal of Separation Science
has provided an alternative way for discrimination of geographic origins or manufacturer and identification of counterfeit drugs. The characterization of counterfeit medicines is based on the identification and quantification of the active substances present. However, in some cases, for example, when low amounts of illegal additives are added, there is no significant difference between counterfeit and genuine drugs. On the other hand, the identification and quantification of active substances require complex pretreatment and more analysis time. Therefore, by omitting the extraction of illegal additives and sample pretreatment, E-nose and flash GC may be powerful techniques for the rapid analysis of counterfeit drugs with the aid of chemometrics. The aim of this work is to establish an approach for rapid discrimination of slimming capsules based on the type difference of illegal additives. E-nose and flash GC were used as two tools for fast analytical technique to obtain the information of the samples. To reveal the difference among the different classes of samples, the principal component analysis (PCA) and linear discriminant analysis (LDA) were employed to analyze the data acquired from E-nose and flash GC, respectively. Moreover, the data fusion strategy was explored to further improve the performance of classifying the slimming capsules according to the difference of illegal additives.
2. Materials and methods 2.1 Materials The slimming capsules were supplied by National Institutes for Food and Drug Control, including twenty unqualified medicines with illegal additives (G1–G3) and two qualified This article is protected by copyright. All rights reserved.
www.jss-journal.com
Page 5
Journal of Separation Science
medicines without illegal additives (G4). The unqualified medicines include three types, which are the samples with phenolphthalein (G1), the samples with sibutramine (G2), and the samples with both phenolphthalein and sibutramine (G3). Detailed information is listed in Table 1. The samples are divided into calibration and prediction set at random. Six samples of G1, four samples of G2 and G3 are used as calibration set. The remaining samples of G1, G2 and G3 are taken as prediction set. Two samples of G4 were used as both calibration and prediction set.
2.2 E-nose analysis -FOX4000 (Alpha M.O.S., Toulouse, France), which consists of an array of sensors, HS-100 autosampler and processing system (Alpha Soft), was used. 18 commercial metal oxide sensors are divided into chambers as three types including T, P and LY sensors. T and P sensors are based on tin dioxide SnO2 (n-type semiconductor), the difference between them resides in the geometry of the sensors. LY sensors are based on chromium titanium oxide (p-type semiconductor) and tungsten oxide (n-type semiconductor). The sensors are mainly used for the analysis of oxidizing gas, ammonia, organic solvents, hydrocarbons, aromatic compounds etc. The samples were accurately weighed to 0.20 g and placed in a 10 mL glass jar, sealed and loaded into the autosampler tray. Each sample was incubated in the bottle at the temperature of 80°C for 600 s under agitation (500 rpm). Then 3.0 mL of the headspace air was automatically injected into the testing chamber by a syringe. The acquisition time was This article is protected by copyright. All rights reserved.
www.jss-journal.com
Page 6
Journal of Separation Science
120 s, a sufficient time for sensors to obtain stable values. The response values of the 18 sensors for every sample were recorded, with response curves generated. When the measurement finished, the clean phase was activated and lasted 360 s, to clean the test chamber and return the sensors to baselines. Useful information may be lost if the data of a sensor during a time period or the data of different sensors at a time were used for the analysis. Therefore, to exploit comprehensive data, the data matrix arranged with 18 arrays of different sensors was used in this work. Fig. 1 displays the measured responses of the sensor with samples. Each curve represents the conductivity of sensor, when volatiles reach the measurement chamber. It can be seen that all the curves are in a high similarity and characteristic response of the samples of four classes cannot be found.
2.3 Flash GC analysis A HERACLES II flash GC electronic nose (Alpha M.O.S, Toulouse, France) was applied to monitor the volatile compound extracting from different samples. This instrument was integrated with classical GC functionalities and E-Nose olfactive fingerprint software. It consisted of a sampling system, a detector system and a data acquisition and processing system (Alpha Soft). The detector system contains two short columns of different polarity and each one coupled to a different flame ionization detector. The same preparation of sample was performed as E-nose analysis. The accumulated gas in the headspace was then injected into GC with 10 m × 0.18 mm MXT-5/MXT-1700 (Alpha M.O.S, Toulouse, France). The carrier gas, hydrogen, was circulated at 1 mL min-1 in the This article is protected by copyright. All rights reserved.
www.jss-journal.com
Page 7
Journal of Separation Science
constant flow mode. The injected volume was 1.0 mL and the injector temperature 200°C. The oven temperature program was as follows: 40°C for 3 s, then 4°C s-1 ramps to 270°C and holding for 30 s. The two flame ionization detectors temperature was 280°C. Fig. 2 displays the measured flash GC curves of two chromatographic columns with samples. It can be seen that all the curves are in a high similarity and seriously overlapped, and characteristic peak of samples of each class cannot be found.
2.4 Discrimination Methods PCA [20,21] is an effective data mining technique and has been widely used for extracting information and reducing data dimensionality. It constructs a new set of orthogonal variables named as principal components (PCs) to describe the original data. The PCs are sorted in a descending order according to the variance explained. Hence, the first few PCs explain most of the variability of the data. The first two or three low-order PCs are generally used for inspection of the classification. If the information of multiple PCs is needed, further operations must be employed for visualization of the classification. In this sturdy, PCA is employed to reduce the dimensions of data. The first two PCs are used for inspection of the classification. LDA [22,23] is one of the most popular techniques of data classification and dimensionality reduction. This method maximizes the variance among categories and minimizes the variance within categories by discriminant functions. It looks for a sensible rule to discriminate categories by forming linear functions of data, and by maximizing the ratio of the between-group sum of squares to the within-group sum of squares. An important This article is protected by copyright. All rights reserved.
www.jss-journal.com
Page 8
Journal of Separation Science
restriction in the application of LDA is that number of samples should be at least three to five times the number of variables. In this study, LDA as a supervised method is investigated to the classification of the samples. Because the number of variables in the spectra is much bigger than the number of samples, LDA will be invalid if the raw data are used as the input data directly. Therefore, PCA is used to reduce the dimension of the spectra and the scores of PCs are used as the input data of LDA.
3. Results and discussion 3.1 Discrimination using PCA PCA is applied to investigate the classification of the E-nose data. Fig. 1 shows the E-nose curves of the samples. It is obvious that the intensity of the sensors vary quite a bit. To eliminate the effect of intensity difference between sensors and make each sensor have a comparable contribution to the classification, normalized data is used in the calculations for the E-nose data. To simplify the inspection, 2-D plot of PC1–PC2 is plotted in the Fig 3 (a). The first two PCs explain 86% of the variance due to the similarity of the sensor response. It can be seen from the figure that, the class of samples without illegal additives (G4) are discriminated in PC1–PC2, although G4 samples is close to the class of samples with both phenolphthalein and sibutramine (G3). However, other classes of samples with different illegal additives are closely overlapped, e.g. samples G1, G2 and G3 are mixed together and cannot be discriminated evidently. PCA is also applied to investigate the classification of the flash GC data. To eliminate the effect of intensity difference between variables, normalized data is also used in the This article is protected by copyright. All rights reserved.
www.jss-journal.com
Page 9
Journal of Separation Science
calculations for the flash GC data. Fig. 3 (b) show that the distribution of samples in the first two PCs by PCA. It can be seen from the figure that, when the samples data are projected into the PC1–PC2 subspaces, the class of slimming medicines without illegal additives is discriminated evidently, e.g. G4 is far from G1, G2 and G3. The other classes, however, are overlapped in PC1–PC2, e.g. G1, G2 and G3 are mixed together. Therefore, the classes of samples with illegal additives cannot be discriminated by using E-nose or flash GC data individually.
3.2 Improving the separability of PCA using data fusion To improve the classification results, the information contained in different instrumental techniques may be explored for classifying such data. Data fusion strategy is applied to obtain a multiple blocks of data information. When multiple blocks of data are collected on the same set of samples, the possibility of integrating the information present in the different matrices into a single model can lead to an improvement of the discrimination. Data fusion [24] strategy consists in concatenating the original data matrices, and then analyzing the resulting array as if it were a single data block. In this study, E-nose data matrix (22 2178) and flash GC data matrix (22 20001) are concatenated a single array (22 22179). Each sample is fingerprinted by E-nose and flash GC instrumental techniques. To eliminate the effect of intensity difference between two instruments and make each instrument have a comparable contribution to the classification, normalized data is used in the calculations. Fig. 3 (c) displays that the scores of samples after fusion data in the first two PCs by PCA. From the figure, it can be seen that G1 and G4 samples are correctly classified in This article is protected by copyright. All rights reserved.
www.jss-journal.com
Page 10
Journal of Separation Science
PC1–PC2 subspace. Clearly, the results are much better than that obtained by one instrument. G2 and G3, however, are still merged by PCA. Therefore, other classification method must be explored for classifying such samples.
3.3 Discrimination using LDA To divide samples into specific groups, LDA, a supervised method, is investigated. In the calculations, calibration samples are used for build the model, and two validation samples selected randomly are used for external validation. According to training set samples, confidence ellipse can be obtained to use for discrimination in classes. The samples can be easily classified as the corresponding class when their scatters are located in the ellipse. Fig. 4 (a) depicts the scatter plot for different samples. In the figure, the samples of G1, G2, G3 and G4 are labeled in blue, red, green and orange, respectively. Moreover, the pentagrams are adopted to label the validation samples. It can be seen from the figure that, G4 samples are distinguished clearly using E-nose data, but samples of G1, G2 and G3 are mixed together. Therefore, classification only using the E-nose data may not be feasible due to the lack more information. Fig. 4 (b) depicts the LDA scatter plot for different samples using flash GC data. It can be seen from the figure that, G4 samples are also distinguished clearly. Although G1 and G2 are divided, but G1 and G2 samples are very close to each other. G3 samples are merged with G1 and G2, respectively. Moreover, confidence ellipses of G1 and G3 have a small part of overlapping, but confidence ellipses of G2 and G3 are seriously overlapping and one sample in G3 is out of the corresponding confidence ellipse. Although classification using the flash This article is protected by copyright. All rights reserved.
www.jss-journal.com
Page 11
Journal of Separation Science
GC data may acquire a better result than using E-nose data, much better performance of classify are needed to be investigated.
3.4 Improving the separability of LDA using data fusion To improve the classification results, LDA is applied for fusion data. Fig. 4 (c) shows the scatter plot after data fusion for different samples by LDA. It is clear that satisfied results are obtained for all the groups, both for the calibration samples and the validation samples labeled with pentagrams. It can be seen from the figure that, G4 is also distinguished clearly, and furthermore all samples are in the corresponding confidence ellipses. On the other hand, the results in Fig 4 (c) are much better than that in Fig. 4 (b). In Fig. 4 (c), group G1 and G2 are divided clearly. Moreover, confidence ellipses of G3 with G2 and G3 only have a small overlap. As a consequence, classification of slimming capsules using fusion data may acquire even more satisfied results than only using E-nose data or flash GC data.
4. Concluding remarks Discrimination of slimming capsules by electronic nose and flash GC coupled to PCA and LDA is studied. The results show that, samples without illegal additives can be discriminated from ones with illegal additives. However, it is difficult to classify the samples from different types illegal only using E-nose or flash GC data. Acceptable classification can be achieved with the help of data fusion between E-nose and flash GC data. Therefore, by omitting the extraction of illegal additives and sample pretreatment, E-nose and flash GC may be powerful techniques for the rapid analysis of slimming capsules with the aid of This article is protected by copyright. All rights reserved.
www.jss-journal.com
Page 12
Journal of Separation Science
chemometrics.
This study was supported by the National Natural Science Foundation of China (No. 21175074).
The authors have declared no conflict of interest.
5. References [1] Youssef R.M., Khamis E.F., Korany M.A., Mahgoub H., Kamal M.F., Anal. Methods 2014, 6, 3395–3400. [2] De Carvalho L.M., Martini M., Moreira A.P.L., de Lima A.P.S., Correia D., Falcao T., Garcia S.C., De Bairros A.V., Do Nascimento P.C., Bohrer D., Forensic Sci. Int. 2010, 204, 6–12. [3] Ariburnu E., Uludag M. F., Yalcinkaya H., Yesilada E., J. Pharmaceut. Biomed. Anal. 2012, 64–65, 77–81. [4] Yu Z., Wei Q., Fan Q., Wan C., J. Liq. Chromatogr. Relat. Technol. 2010, 33, 452–461. [5] Sharma K., Sharma S.P., Lahiri S.C., J. Forensic Sci. 2013, S1, S208-S214. [6] Luque C.A., Rey J.A., Eur. J. Pharmacol. 2002, 440, 119–128. [7] Degardin K., Roggo Y., Margot P., J. Pharmaceut. Biomed. Anal. 2014, 87, 167–175. This article is protected by copyright. All rights reserved.
www.jss-journal.com
Page 13
Journal of Separation Science
[8] Fraser S.J., Oughton J., Batten W.A., Clark A.S.S., Schmierer D.M., Gordon K.C., Strachan C.J., J. Raman Spectrosc. 2013, 8 1172–1180. [9]
Custers D., Canfy M., Courselle P., De Beer J.O., Apers S., Deconinck E., Talanta 2014, 12, 78–88.
[10] Baldwin E.A., Bai J., Plotto A., Bea S., Sensors 2011, 11, 4744–4766. [11] Barbri N.E.I, Mirhisse J., Ionescu R., Correig X., Bouchikhi B., Llobet E., Sens. Actuators B: Chem. 2009, 141, 538–543. [12] Apetrei C., Apetrei I., Villanueva S., Saja J., Gutierrez-Rosales F., RodriguezMendez M., Anal. Chim. Acta 2010, 663, 91–97. [13] Blasioli S., Biondi E., Braschi I., Mazzucchi U., Bazzi C., Gessa C., Anal. Chim. Acta 2010, 672, 20–24. [14] Xiong Y., Xiao X.H., Yang X.Y., Yan D., Zhang C., Zou H.Q., Lin H., Peng L., Xiao X., Yan Y.H., J. Pharmaceut. Biomed. Anal. 2014, 91, 68–72. [15] Li C., Xu F., Cao C., Shang M.Y., Zhang C.Y., Yu J., Liu G.X., Wang X., Cai S.Q., J. Pharmaceut. Biomed. Anal. 2013, 85, 231–238. [16] Xiao Z.B., Yu D., Niu Y.W., Chen F., Song S.Q., Zhu J.C., Zhu G.Y., J. Chromatogr. B 2014, 945–946, 92–100. [17] Zhang M.L., Harrington P. D. B., Talanta 2013, 117, 483–491. [18] Binette M.J., Lafontaine P., Vanier M., J. Agric. Food Chem. 2009, 57, 1151–1155. [19] Delahunty C.M., Eyres G., Dufour J.P., J. Sep. Sci. 2006, 29, 2107–2125. [20] Joliffe I. T., Principal Component Analysis, Springer, New York, 1986.
This article is protected by copyright. All rights reserved.
www.jss-journal.com
Page 14
Journal of Separation Science
[21] Costache G.N., Corcoran P., Puslecki P., Pattern Recogn. Lett. 2009, 30, 1441–1447. [22] Tominaga Y., Chemometr. Intell. Lab. 1999, 49, 105–115. [23] Lin H., Yan Y.H., Zhao T., Peng L., Zou H.Q., Yang X.Y., Xiong Y., Wang M., Wu H.Z., J. Pharmaceut. Biomed. Anal. 2013, 84, 1–4. [24] Biancolillo A., Bucci R., Magri A.L., Magri A.D., Marini F., Anal. Chim. Acta 2014, 820, 23–31.
This article is protected by copyright. All rights reserved.
www.jss-journal.com
Page 15
Fig. 1 Measured response curves of 18 sensors for the samples.
This article is protected by copyright. All rights reserved.
Journal of Separation Science
www.jss-journal.com
Page 16
Fig. 2 Measured flash GC profiles of two columns for the samples.
This article is protected by copyright. All rights reserved.
Journal of Separation Science
www.jss-journal.com
Page 17
Journal of Separation Science
Fig. 3 Distribution of the samples in PC1–PC2 using E-nose data (a), flash GC data (b) and fused data (c).
This article is protected by copyright. All rights reserved.
www.jss-journal.com
Page 18
Journal of Separation Science
Fig. 4 Linear discrimination analysis (LDA) scatter plots for the samples using E-nose data (a), flash GC data (b) and fused data (c).
This article is protected by copyright. All rights reserved.
www.jss-journal.com
Page 19
Journal of Separation Science
Table 1 Detail information of the slimming capsule samples.
Medicine No. 1-8 9-14 15-20
Class Label G1 G2 G3
21-22
G4
This article is protected by copyright. All rights reserved.
Illegal Additives Phenolphthalein Sibutramine Phenolphthalein and sibutramine Not added
Page 1
Journal of Separation Science
Rapid discrimination of slimming capsules based on illegal additives by electronic nose and flash gas chromatography
Zhenzhen Xia, Wensheng Cai, Xueguang Shao Collaborative Innovation Center of Chemical Science and Engineering (Tianjin), State Key Laboratory of Medicinal Chemical Biology, and Research Center for Analytical Sciences, College of Chemistry, Nankai University, Tianjin, 300071, China
Keywords: Counterfeit medicines / Data fusion / Electronic nose / Flash gas chromatography / Linear discriminant analysis
Correspondence: Pro. Xueguang Shao, College of Chemistry, Nankai University, Tianjin, 300071, P. R. China Fax: +86-22-23502458 E-mail: [email protected]
Abbreviations: E-nose, electronic nose; GC, gas chromatography; PCA, principal component analysis; LDA, linear discriminant analysis; Received: 25-Aug-2014; Revised: 07-Nov-2014; Accepted: 19-Nov-2014 This article has been accepted for publication and undergone full peer review but has not been through the copyediting, typesetting, pagination and proofreading process, which may lead to differences between this version and the Version of Record. Please cite this article as doi: 10.1002/jssc.201400941. This article is protected by copyright. All rights reserved.
www.jss-journal.com
Page 2
Journal of Separation Science
Abstract The discrimination of counterfeit and/or illegally manufactured medicines is an important task in the pharmaceutical industry for pharmaceutical safety. In this study, 22 slimming capsule samples with illegally added sibutramine and phenolphthalein were analyzed by electronic nose and flash gas chromatography. To reveal the difference among the different classes of samples, the principal component analysis and linear discriminant analysis were employed to analyze the data acquired from electronic nose and flash gas chromatography, respectively. The samples without illegal additives can be discriminated from the ones with illegal additives by using electronic nose or flash gas chromatography data individually. To improve the performance of classification, data fusion strategy was applied to integrate the data from electronic nose and flash gas chromatography data into a single model. The results show that the samples with phenolphthalein, sibutramine and both can be classified well by using fused data.
1. Introduction Excess weight and obesity have been recognized as serious worldwide problems leading to the search for a number of weight-loss medicines, especially alternative treatments based on natural slimming formulations [1]. Although natural slimming formulations present some side effects, it is a fact that the potency of these products is lower than synthetic slimming agents [2,3]. As a result, some manufacturers committed illegal adulterations of synthetic anorexics and cathartics to improve the efficacy of their products. Phenolphthalein and sibutramine are often added illegally into natural slimming products [4]. Phenolphthalein is This article is protected by copyright. All rights reserved.
www.jss-journal.com
Page 3
Journal of Separation Science
an acid–base indicator and laxative, which is prescribed for short-term treatment of constipation in humans, implying its possible weight-reducing effect. However, it can cause more severe constipation, electrolyte imbalance, cancer, cardiac arrhythmia and even death [5]. Sibutramine is a serotonin noradrenalin reuptake inhibitor, which promotes and maintains weight loss in obese people. However, it simultaneously overexcites the central nervous system and several side effects like nervousness, xerostomia, headache, numbness and paraesthesia have been reported [6]. Counterfeit medicines with illegal additives pose a huge threat to public health worldwide. Hence, the discrimination of counterfeit medicines has been an important task in pharmaceutical industry [7,8]. Odors of natural slimming medicine could vary between genuine and counterfeit medicines [9]. Therefore, to detect the subtle differences of odors is a feasible way to distinguish natural slimming medicine. The electronic nose (E-nose) [10], engineered to mimic the mammalian olfactory system, is designed to acquire objective data and perform odor evaluation on a continuous basis with a minimal cost. It identifies the mixture of volatile constituents as a whole without having to detect the individual chemical constituents within the mixture. Its array of cross-reactive sensors can interact with a broad range of volatile compounds, and the multivariate responses can be used as an “electronic fingerprint” to identify the odors through pattern recognition. Its advantages also include no requirement of sample pretreatment and rapid and sensitive analysis [11–13]. Electronic noses have been applied successfully in natural medicine analysis [14]. Flash GC also offers an effective and rapid tool for the volatiles characterization and detection in the different area [15–19], like pharmaceutical analysis, food analysis, environmental analysis, etc. Besides, chemometrics This article is protected by copyright. All rights reserved.
www.jss-journal.com
Page 4
Journal of Separation Science
has provided an alternative way for discrimination of geographic origins or manufacturer and identification of counterfeit drugs. The characterization of counterfeit medicines is based on the identification and quantification of the active substances present. However, in some cases, for example, when low amounts of illegal additives are added, there is no significant difference between counterfeit and genuine drugs. On the other hand, the identification and quantification of active substances require complex pretreatment and more analysis time. Therefore, by omitting the extraction of illegal additives and sample pretreatment, E-nose and flash GC may be powerful techniques for the rapid analysis of counterfeit drugs with the aid of chemometrics. The aim of this work is to establish an approach for rapid discrimination of slimming capsules based on the type difference of illegal additives. E-nose and flash GC were used as two tools for fast analytical technique to obtain the information of the samples. To reveal the difference among the different classes of samples, the principal component analysis (PCA) and linear discriminant analysis (LDA) were employed to analyze the data acquired from E-nose and flash GC, respectively. Moreover, the data fusion strategy was explored to further improve the performance of classifying the slimming capsules according to the difference of illegal additives.
2. Materials and methods 2.1 Materials The slimming capsules were supplied by National Institutes for Food and Drug Control, including twenty unqualified medicines with illegal additives (G1–G3) and two qualified This article is protected by copyright. All rights reserved.
www.jss-journal.com
Page 5
Journal of Separation Science
medicines without illegal additives (G4). The unqualified medicines include three types, which are the samples with phenolphthalein (G1), the samples with sibutramine (G2), and the samples with both phenolphthalein and sibutramine (G3). Detailed information is listed in Table 1. The samples are divided into calibration and prediction set at random. Six samples of G1, four samples of G2 and G3 are used as calibration set. The remaining samples of G1, G2 and G3 are taken as prediction set. Two samples of G4 were used as both calibration and prediction set.
2.2 E-nose analysis -FOX4000 (Alpha M.O.S., Toulouse, France), which consists of an array of sensors, HS-100 autosampler and processing system (Alpha Soft), was used. 18 commercial metal oxide sensors are divided into chambers as three types including T, P and LY sensors. T and P sensors are based on tin dioxide SnO2 (n-type semiconductor), the difference between them resides in the geometry of the sensors. LY sensors are based on chromium titanium oxide (p-type semiconductor) and tungsten oxide (n-type semiconductor). The sensors are mainly used for the analysis of oxidizing gas, ammonia, organic solvents, hydrocarbons, aromatic compounds etc. The samples were accurately weighed to 0.20 g and placed in a 10 mL glass jar, sealed and loaded into the autosampler tray. Each sample was incubated in the bottle at the temperature of 80°C for 600 s under agitation (500 rpm). Then 3.0 mL of the headspace air was automatically injected into the testing chamber by a syringe. The acquisition time was This article is protected by copyright. All rights reserved.
www.jss-journal.com
Page 6
Journal of Separation Science
120 s, a sufficient time for sensors to obtain stable values. The response values of the 18 sensors for every sample were recorded, with response curves generated. When the measurement finished, the clean phase was activated and lasted 360 s, to clean the test chamber and return the sensors to baselines. Useful information may be lost if the data of a sensor during a time period or the data of different sensors at a time were used for the analysis. Therefore, to exploit comprehensive data, the data matrix arranged with 18 arrays of different sensors was used in this work. Fig. 1 displays the measured responses of the sensor with samples. Each curve represents the conductivity of sensor, when volatiles reach the measurement chamber. It can be seen that all the curves are in a high similarity and characteristic response of the samples of four classes cannot be found.
2.3 Flash GC analysis A HERACLES II flash GC electronic nose (Alpha M.O.S, Toulouse, France) was applied to monitor the volatile compound extracting from different samples. This instrument was integrated with classical GC functionalities and E-Nose olfactive fingerprint software. It consisted of a sampling system, a detector system and a data acquisition and processing system (Alpha Soft). The detector system contains two short columns of different polarity and each one coupled to a different flame ionization detector. The same preparation of sample was performed as E-nose analysis. The accumulated gas in the headspace was then injected into GC with 10 m × 0.18 mm MXT-5/MXT-1700 (Alpha M.O.S, Toulouse, France). The carrier gas, hydrogen, was circulated at 1 mL min-1 in the This article is protected by copyright. All rights reserved.
www.jss-journal.com
Page 7
Journal of Separation Science
constant flow mode. The injected volume was 1.0 mL and the injector temperature 200°C. The oven temperature program was as follows: 40°C for 3 s, then 4°C s-1 ramps to 270°C and holding for 30 s. The two flame ionization detectors temperature was 280°C. Fig. 2 displays the measured flash GC curves of two chromatographic columns with samples. It can be seen that all the curves are in a high similarity and seriously overlapped, and characteristic peak of samples of each class cannot be found.
2.4 Discrimination Methods PCA [20,21] is an effective data mining technique and has been widely used for extracting information and reducing data dimensionality. It constructs a new set of orthogonal variables named as principal components (PCs) to describe the original data. The PCs are sorted in a descending order according to the variance explained. Hence, the first few PCs explain most of the variability of the data. The first two or three low-order PCs are generally used for inspection of the classification. If the information of multiple PCs is needed, further operations must be employed for visualization of the classification. In this sturdy, PCA is employed to reduce the dimensions of data. The first two PCs are used for inspection of the classification. LDA [22,23] is one of the most popular techniques of data classification and dimensionality reduction. This method maximizes the variance among categories and minimizes the variance within categories by discriminant functions. It looks for a sensible rule to discriminate categories by forming linear functions of data, and by maximizing the ratio of the between-group sum of squares to the within-group sum of squares. An important This article is protected by copyright. All rights reserved.
www.jss-journal.com
Page 8
Journal of Separation Science
restriction in the application of LDA is that number of samples should be at least three to five times the number of variables. In this study, LDA as a supervised method is investigated to the classification of the samples. Because the number of variables in the spectra is much bigger than the number of samples, LDA will be invalid if the raw data are used as the input data directly. Therefore, PCA is used to reduce the dimension of the spectra and the scores of PCs are used as the input data of LDA.
3. Results and discussion 3.1 Discrimination using PCA PCA is applied to investigate the classification of the E-nose data. Fig. 1 shows the E-nose curves of the samples. It is obvious that the intensity of the sensors vary quite a bit. To eliminate the effect of intensity difference between sensors and make each sensor have a comparable contribution to the classification, normalized data is used in the calculations for the E-nose data. To simplify the inspection, 2-D plot of PC1–PC2 is plotted in the Fig 3 (a). The first two PCs explain 86% of the variance due to the similarity of the sensor response. It can be seen from the figure that, the class of samples without illegal additives (G4) are discriminated in PC1–PC2, although G4 samples is close to the class of samples with both phenolphthalein and sibutramine (G3). However, other classes of samples with different illegal additives are closely overlapped, e.g. samples G1, G2 and G3 are mixed together and cannot be discriminated evidently. PCA is also applied to investigate the classification of the flash GC data. To eliminate the effect of intensity difference between variables, normalized data is also used in the This article is protected by copyright. All rights reserved.
www.jss-journal.com
Page 9
Journal of Separation Science
calculations for the flash GC data. Fig. 3 (b) show that the distribution of samples in the first two PCs by PCA. It can be seen from the figure that, when the samples data are projected into the PC1–PC2 subspaces, the class of slimming medicines without illegal additives is discriminated evidently, e.g. G4 is far from G1, G2 and G3. The other classes, however, are overlapped in PC1–PC2, e.g. G1, G2 and G3 are mixed together. Therefore, the classes of samples with illegal additives cannot be discriminated by using E-nose or flash GC data individually.
3.2 Improving the separability of PCA using data fusion To improve the classification results, the information contained in different instrumental techniques may be explored for classifying such data. Data fusion strategy is applied to obtain a multiple blocks of data information. When multiple blocks of data are collected on the same set of samples, the possibility of integrating the information present in the different matrices into a single model can lead to an improvement of the discrimination. Data fusion [24] strategy consists in concatenating the original data matrices, and then analyzing the resulting array as if it were a single data block. In this study, E-nose data matrix (22 2178) and flash GC data matrix (22 20001) are concatenated a single array (22 22179). Each sample is fingerprinted by E-nose and flash GC instrumental techniques. To eliminate the effect of intensity difference between two instruments and make each instrument have a comparable contribution to the classification, normalized data is used in the calculations. Fig. 3 (c) displays that the scores of samples after fusion data in the first two PCs by PCA. From the figure, it can be seen that G1 and G4 samples are correctly classified in This article is protected by copyright. All rights reserved.
www.jss-journal.com
Page 10
Journal of Separation Science
PC1–PC2 subspace. Clearly, the results are much better than that obtained by one instrument. G2 and G3, however, are still merged by PCA. Therefore, other classification method must be explored for classifying such samples.
3.3 Discrimination using LDA To divide samples into specific groups, LDA, a supervised method, is investigated. In the calculations, calibration samples are used for build the model, and two validation samples selected randomly are used for external validation. According to training set samples, confidence ellipse can be obtained to use for discrimination in classes. The samples can be easily classified as the corresponding class when their scatters are located in the ellipse. Fig. 4 (a) depicts the scatter plot for different samples. In the figure, the samples of G1, G2, G3 and G4 are labeled in blue, red, green and orange, respectively. Moreover, the pentagrams are adopted to label the validation samples. It can be seen from the figure that, G4 samples are distinguished clearly using E-nose data, but samples of G1, G2 and G3 are mixed together. Therefore, classification only using the E-nose data may not be feasible due to the lack more information. Fig. 4 (b) depicts the LDA scatter plot for different samples using flash GC data. It can be seen from the figure that, G4 samples are also distinguished clearly. Although G1 and G2 are divided, but G1 and G2 samples are very close to each other. G3 samples are merged with G1 and G2, respectively. Moreover, confidence ellipses of G1 and G3 have a small part of overlapping, but confidence ellipses of G2 and G3 are seriously overlapping and one sample in G3 is out of the corresponding confidence ellipse. Although classification using the flash This article is protected by copyright. All rights reserved.
www.jss-journal.com
Page 11
Journal of Separation Science
GC data may acquire a better result than using E-nose data, much better performance of classify are needed to be investigated.
3.4 Improving the separability of LDA using data fusion To improve the classification results, LDA is applied for fusion data. Fig. 4 (c) shows the scatter plot after data fusion for different samples by LDA. It is clear that satisfied results are obtained for all the groups, both for the calibration samples and the validation samples labeled with pentagrams. It can be seen from the figure that, G4 is also distinguished clearly, and furthermore all samples are in the corresponding confidence ellipses. On the other hand, the results in Fig 4 (c) are much better than that in Fig. 4 (b). In Fig. 4 (c), group G1 and G2 are divided clearly. Moreover, confidence ellipses of G3 with G2 and G3 only have a small overlap. As a consequence, classification of slimming capsules using fusion data may acquire even more satisfied results than only using E-nose data or flash GC data.
4. Concluding remarks Discrimination of slimming capsules by electronic nose and flash GC coupled to PCA and LDA is studied. The results show that, samples without illegal additives can be discriminated from ones with illegal additives. However, it is difficult to classify the samples from different types illegal only using E-nose or flash GC data. Acceptable classification can be achieved with the help of data fusion between E-nose and flash GC data. Therefore, by omitting the extraction of illegal additives and sample pretreatment, E-nose and flash GC may be powerful techniques for the rapid analysis of slimming capsules with the aid of This article is protected by copyright. All rights reserved.
www.jss-journal.com
Page 12
Journal of Separation Science
chemometrics.
This study was supported by the National Natural Science Foundation of China (No. 21175074).
The authors have declared no conflict of interest.
5. References [1] Youssef R.M., Khamis E.F., Korany M.A., Mahgoub H., Kamal M.F., Anal. Methods 2014, 6, 3395–3400. [2] De Carvalho L.M., Martini M., Moreira A.P.L., de Lima A.P.S., Correia D., Falcao T., Garcia S.C., De Bairros A.V., Do Nascimento P.C., Bohrer D., Forensic Sci. Int. 2010, 204, 6–12. [3] Ariburnu E., Uludag M. F., Yalcinkaya H., Yesilada E., J. Pharmaceut. Biomed. Anal. 2012, 64–65, 77–81. [4] Yu Z., Wei Q., Fan Q., Wan C., J. Liq. Chromatogr. Relat. Technol. 2010, 33, 452–461. [5] Sharma K., Sharma S.P., Lahiri S.C., J. Forensic Sci. 2013, S1, S208-S214. [6] Luque C.A., Rey J.A., Eur. J. Pharmacol. 2002, 440, 119–128. [7] Degardin K., Roggo Y., Margot P., J. Pharmaceut. Biomed. Anal. 2014, 87, 167–175. This article is protected by copyright. All rights reserved.
www.jss-journal.com
Page 13
Journal of Separation Science
[8] Fraser S.J., Oughton J., Batten W.A., Clark A.S.S., Schmierer D.M., Gordon K.C., Strachan C.J., J. Raman Spectrosc. 2013, 8 1172–1180. [9]
Custers D., Canfy M., Courselle P., De Beer J.O., Apers S., Deconinck E., Talanta 2014, 12, 78–88.
[10] Baldwin E.A., Bai J., Plotto A., Bea S., Sensors 2011, 11, 4744–4766. [11] Barbri N.E.I, Mirhisse J., Ionescu R., Correig X., Bouchikhi B., Llobet E., Sens. Actuators B: Chem. 2009, 141, 538–543. [12] Apetrei C., Apetrei I., Villanueva S., Saja J., Gutierrez-Rosales F., RodriguezMendez M., Anal. Chim. Acta 2010, 663, 91–97. [13] Blasioli S., Biondi E., Braschi I., Mazzucchi U., Bazzi C., Gessa C., Anal. Chim. Acta 2010, 672, 20–24. [14] Xiong Y., Xiao X.H., Yang X.Y., Yan D., Zhang C., Zou H.Q., Lin H., Peng L., Xiao X., Yan Y.H., J. Pharmaceut. Biomed. Anal. 2014, 91, 68–72. [15] Li C., Xu F., Cao C., Shang M.Y., Zhang C.Y., Yu J., Liu G.X., Wang X., Cai S.Q., J. Pharmaceut. Biomed. Anal. 2013, 85, 231–238. [16] Xiao Z.B., Yu D., Niu Y.W., Chen F., Song S.Q., Zhu J.C., Zhu G.Y., J. Chromatogr. B 2014, 945–946, 92–100. [17] Zhang M.L., Harrington P. D. B., Talanta 2013, 117, 483–491. [18] Binette M.J., Lafontaine P., Vanier M., J. Agric. Food Chem. 2009, 57, 1151–1155. [19] Delahunty C.M., Eyres G., Dufour J.P., J. Sep. Sci. 2006, 29, 2107–2125. [20] Joliffe I. T., Principal Component Analysis, Springer, New York, 1986.
This article is protected by copyright. All rights reserved.
www.jss-journal.com
Page 14
Journal of Separation Science
[21] Costache G.N., Corcoran P., Puslecki P., Pattern Recogn. Lett. 2009, 30, 1441–1447. [22] Tominaga Y., Chemometr. Intell. Lab. 1999, 49, 105–115. [23] Lin H., Yan Y.H., Zhao T., Peng L., Zou H.Q., Yang X.Y., Xiong Y., Wang M., Wu H.Z., J. Pharmaceut. Biomed. Anal. 2013, 84, 1–4. [24] Biancolillo A., Bucci R., Magri A.L., Magri A.D., Marini F., Anal. Chim. Acta 2014, 820, 23–31.
This article is protected by copyright. All rights reserved.
www.jss-journal.com
Page 15
Fig. 1 Measured response curves of 18 sensors for the samples.
This article is protected by copyright. All rights reserved.
Journal of Separation Science
www.jss-journal.com
Page 16
Fig. 2 Measured flash GC profiles of two columns for the samples.
This article is protected by copyright. All rights reserved.
Journal of Separation Science
www.jss-journal.com
Page 17
Journal of Separation Science
Fig. 3 Distribution of the samples in PC1–PC2 using E-nose data (a), flash GC data (b) and fused data (c).
This article is protected by copyright. All rights reserved.
www.jss-journal.com
Page 18
Journal of Separation Science
Fig. 4 Linear discrimination analysis (LDA) scatter plots for the samples using E-nose data (a), flash GC data (b) and fused data (c).
This article is protected by copyright. All rights reserved.
www.jss-journal.com
Page 19
Journal of Separation Science
Table 1 Detail information of the slimming capsule samples.
Medicine No. 1-8 9-14 15-20
Class Label G1 G2 G3
21-22
G4
This article is protected by copyright. All rights reserved.
Illegal Additives Phenolphthalein Sibutramine Phenolphthalein and sibutramine Not added
