Fluorescence Spectral Analysis for the Discrimination of Complex,Similar Mixtures with the Aid of Chemometrics

Abstract

An analytical method for the classification of complex real-world samples was researched and developed with the use of excitation-emission fluorescence matrix (EEFM) spectroscopy, using the medicinal herbs, Rhizoma corydalis decumbentis (RCD) and Rhizoma corydalis (RC) as example samples. The data set was obtained from various authentic RCD-A and RC-A, adulterated AD, and commercial RCD-C and RC-C samples. The spectra (range: λ_ex=215∼395 nm and λ_em=290∼560 nm), arranged in two- and three-way data matrix formats, were processed using principal component analysis (PCA) and parallel factor analysis (PARAFAC) to produce two-dimensional component-by-component plots for qualitative data classification. The RCD-A and RC-A object groups were clearly discriminated, but the AD and the RCD-C as well as RC-C samples were less well separated. PARAFAC analysis produced somewhat better discrimination, and loadings plots revealed the presence of the marker compound Protopine—a strongly fluorescing substance—as well as at least two other unidentified fluorescent components. Classification performance of the common K-nearest neighbors (KNN) and linear discrimination analysis (LDA) methods was relatively poor when compared with that of the back propagation- and radial basis function-artificial neural networks (BP-ANN and RBF-ANN) models on the basis of two- and three-way formatted data. The best results were obtained with the three-way fingerprints and the RBF-ANN model. Subsequently, the quality of the commercial samples (RCD-C and RC-C) was classified on the best optimized RBF-ANN model. Thus, EEFM spectroscopy, which provides three-way measured data, is potentially a powerful analytical technique for the analysis of complex real-world substances provided the classification is performed by the RBF-ANN or similar ANN methods.

Keywords

Fluorescence spectroscopy Excitation-emission fluorescence Classification Artificial neural networks Principal component analysis PCA Parallel factor analysis PARAFAC

INTRODUCTION

Excitation-emission fluorescence (EEF) spectroscopy is a well-established technique for fingerprint analysis of samples, which, often, contain polycyclic aromatic hydrocarbons. Examples of such applications include the identification and classification of various kinds of oils,^1–4 water samples,⁵ pharmaceuticals,⁶ tobacco,⁷ and dyes.⁸ The technique is rapid, has high sensitivity, requires small amounts of sample,⁹ and provides qualitative and quantitative information. Sample preparation is often quite simple, involving just an aqueous or buffer solution extraction, and fluorescence spectroscopy will normally detect the fluorescing substances so extracted (limit of detection (LOD) ∼10⁻⁶–10⁻⁹ mol L⁻¹) irrespective of any other components that may have dissolved in the extraction medium. Useful and convenient substances for research and development of analytical methodologies for comparison of complex samples are herbs because they are complex plant materials and are in common use throughout the world. Fluorescing molecular components of a traditional Chinese medicine (TCM) are often the active species; they are usually aromatic compounds or their derivatives, which have a π-conjugation and a rigid, planar structure, e.g., alkaloids, coumarins, flavonoids, anthraquinones, and phenylpropanoids.

Recently, there has been increasing interest in the development of analytical methodology for comparison of samples containing such substances so as to ensure their quality and to limit the spread of adulterated, substituted, and alternative TCMs. In this context, Wei et al.^10,11 applied excitation-emission fluorescence matrix (EEFM) spectroscopy quantitatively to analyze the contents of such substances but the work did not extend to cover any comparisons, classification, or discrimination amongst such samples. Ortiz et al.¹² found ciprofloxacin in human urine from EEFM data submitted to parallel factor (PARAFAC) analysis, and the associated recovery and the absolute relative errors were 88.3% and 4.2%, respectively. Maggio et al.¹³ presented a new application for the three-way kinetic EEFM data for quantitative determination of two inter-converting analytes, carbaryl (0– 363 μg L⁻¹) and 1-naphthol (0–512 μg L⁻¹), embedded in a complex sample background. Data processing was performed with the unfolded partial least squares (PLS) method combined with residual tri-linearization (U-PLS/RTL), and also with the (PARAFAC) model, combined with a calibration based on multilinear regression. U-PLS/RTL was shown to be significantly simpler, and it produced similar figures of merit in comparison to those from the PARAFAC analysis. The application and algorithms of EEF for second- and third-order calibration can be found in two recent reviews.^14,15

For this work, two similar TCMs were chosen: (i) Rhizoma corydalis decumbentis (RCD), which is the dried root tuber of Corydalis decumbens (Thunb.) Pers., and a plant of the Papaveraceae family, which is widely distributed in the Jiangxi province, and (ii) Rhizoma corydalis (RC), which is the dried root tuber of Corydalis yanhusuo W.T. Wang, of the same family, and mainly grows in Zhejiang province. Both are well-known TCMs with many well-documented medicinal properties.^16–19 Their actual appearance is very similar (Fig. 1A), and as a consequence they can be used to adulterate each other or as substitutes. However, even though some of the plant properties are similar, for clinical purposes, it is desirable to know accurately the nature of a prescribed medication; furthermore, the two TCMs often have different market values depending on their availability at the time. In such cases, adulteration of the more expensive by the cheaper one can occur. This indicates that the above two TCMs are particularly appropriate examples for investigation because it focuses on the development of simple, rapid analytical methodologies for the discrimination and classification of complex substances with the use of fluorescence spectral fingerprints and their interpretation with the use of chemometrics methods.

Fig. 1.

(A) Appearance of Rhizoma corydalis decumbentis (RCD-A) and Rhizoma corydalis (RC-A); (B) PC1 versus PC2 bi-plot for the 119 TCM samples; data, two-way fingerprints.

In general, examples of classical methods for TCM identification include morphological and microscopic examination, both of which are highly subjective, particularly with the very similar RCD and RC root samples (Fig. 1A). Previous phytochemical and clinical studies have shown that protopine and tetrahydropalmatine have been identified as the active secondary metabolites of the plants, and they were noted as reference markers.²⁰ Protopine has different effects^21,22 when interacting with tetrahydropalmatine,^23,24 but both are present in the two TCM plants as well as in other members of the Corydalis family;^25,26 thus, the use of these substances as markers for comparative purposes of RCD and RC is insufficient. In addition, previous literature for RCD and RC mostly focused on the identification and quantification of one or more compounds, such as protopine, tetrahydropalmatine, and palmatine, with the use of liquid chromatography (LC) or capillary electrophoresis (CE).^16,27 Other studies have compared the different Corydalis species,²⁷ but these involved quite expensive mass-spectrometric techniques, which were time consuming and required organic solvents. Recent studies suggested that alkaloids were the main active constituents separated from RCD and RC,¹⁶ and it was demonstrated that the marker compound protopine fluoresces. This suggests that EEFM spectroscopy is suitable for studying both RCD and RC.

The aims of this study were to research and develop a rapid and reliable screening method with the use of EEFM fingerprints to enable the classification of complex, real-world samples such as the two visually similar RCD and RC materials; of particular interest was the performance of the chemometrics classification models to discriminate the adulterated samples. Thus, it was the intention of this work to explore the use of two- and three-way EEFM spectral fingerprints for qualitative analysis with the use of principal component analysis (PCA) and parallel factor analysis (PARAFAC), respectively. Several chemometrics methods: K-nearest neighbors (KNN), linear discrimination analysis (LDA), back propagation artificial neural networks (BP-ANN), and radial basis function artificial neural networks (RBF-ANN) were then used for classification of the samples on the basis of the two-way fingerprint data, and their performance was compared.

EXPERIMENTAL

Materials and Reagents. Sixty RCD and thirty RC samples (RCD-A and RC-A) were obtained from different sources and authenticated (Prof. Rao, Jiangxi University of Traditional Chinese Medicine, China). Fifteen adulterated samples (AD) were prepared as follows: one authentic RCD sample was selected at random, ground, and mixed with a randomly selected authentic RC sample in ratios of 1:4, 4:1, and 1:1. This procedure was repeated five times and a total of 15 AD (AD-1, AD-2, and AD-3) samples were obtained. Fourteen commercial samples were purchased from pharmacies in Jiangxi, including 11 RCD and 3 RC samples (RCD-C and RC-C), which gave a total of 119 analytical samples (Table I). The quality of these commercial samples was uncertain and yet to be evaluated; they serve as unknown samples in this work.

TABLE I.

Test samples information.

Sample name and number^a	Size	(RCD/RC) %	Source
RCD-A (1–12)	12	(100/0) %	Guangxi (original)
RCD-A (13–60)	48	(100/0) %	Jiangxi (original)
AD-1 (61–65)	5	(80/20) %	RCD/RC (4/1; mixed)
AD-2 (66–70)	5	(50/50) %	RCD/RC (1/1; mixed)
AD-3 (71–75)	5	(20/80) %	RCD/RC (1/4; mixed)
RC-A (76–93)	18	(0/100) %	Zhengjiang (original)
RC-A (94–105)	12	(0/100) %	Hebei (original)
RCD-C (106–116)	11	Unknown	Purchased (Jiangxi)
RC-C (117–119)	3	Unknown	Purchased (Jiangxi)

RCD-A and RC-A: authentic Rhizoma corydalis decumbentis (RCD) and Rhizoma corydalis (RC); samples collected from the original growing areas; AD-1,–2, and–3: mixed or adulterated samples of RCD and RC ratios of 4:1, 1:1, and 1:4, respectively; RCD-C and RC-C: commercial Rhizoma corydalis decumbentis and Rhizoma corydalis samples purchased from the TCM markets (Jiangxi; samples' quality uncertain and to be evaluated; used in this study as unknown samples).

Protopine (Scheme 1A) and tetrahydropalmatine (Scheme 1B) were obtained from the National Institute for the Control of Pharmacological and Biological Products of China, Beijing, China. Freshly redistilled water was used throughout the experiments. Protopine solution (0.25 mg mL⁻¹) was prepared by dissolving 6.25 mg crystal in 25 mL methanol.

Scheme 1.

Molecular structure of (A) protopine and (B) tetrahydropalmatine.

Sample Preparation. The 119 samples were ground into powder and passed through a 60-mesh sieve. Each powdered sample (0.10 g) was accurately weighed and extracted for 30 min with redistilled water (20.0 mL) with the use of an ultrasonic water-bath (60 ± 0.5 °C). Then, the sample was filtered and washed with 50 mL water, after which the collected solvent was diluted to 100 mL with water.

Instrumentation. Each of the prepared samples was diluted 150 times and measured on a Perkin-Elmer LS-55 spectrofluorimeter equipped with a thermostatic bath (Model ZC-10, Ningbo Tianheng Instruments Factory, China), a xenon lamp, and a 10 mm quartz cuvette. Emission spectra were collected in the range of 290 to 560 nm in 0.5 nm steps (total, 541 data points), while the excitation wavelengths were scanned from 215 to 395 nm every 5 nm (total, 37 data points), to produce a 541 (emission wavelengths) × 37 (excitation wavelengths) EEFM matrix. The bandwidths of both monochromators (5 nm resolution) and the scan rate (1500 nm min⁻¹) were held constant for all spectral measurements. An emission filter (290 nm) was used to eliminate the second-order Rayleigh scattering from the excitation monochromator. Note that the EEFM of pure protopine was also measured under the same conditions. The FL-Winlab Software (Perkin Elmer) was used for collecting the measured data. All experiments were carried out at a temperature of 25 ± 1 °C.

Development of Two-Way Fingerprints Based on EEFM Data Using PCA. Suppose X (N×M) is a two-dimensional data matrix of each sample measured by EEF with N emission wavelengths and M excitation wavelengths. X can be described as M one-dimensional emission spectra at various excitation wavelengths, and each emission spectrum is measured at N regular wavelength intervals, as described below:

\begin{array}{l} \begin{array}{l} X = [x_{1}, x_{2}, ..., x_{M}] \\ x_{j} = [x_{1 i}, x_{2 i}, ..., x_{N i}] & (i = 1, 2, ..., M; j = 1, 2, ... N) \end{array} \end{array}

(1)

where the vector x_j (N × 1) denotes the emission spectrum measured at the ith excitation wavelength. By application of PCA,^28,29 X can be decomposed into three matrices, the score matrix T (N×M), the loadings matrix V^T (M×M), and the residual matrix E (N×M):

X = T \times V^{T} + E

(2)

The score matrix, T, captures the systematic variation in the data pertaining to the objects, and the first score, t₁ (N × 1), which contains most of the significant information, is extracted. If the data matrix of P tested samples is obtained with the use of EEFM, i.e., O (P×N×M), a group of scores, t₁, each from a different sample, may be extracted; then a new data matrix with P samples and N variables may be produced by PCA. This new data matrix was denoted as O^new (P × N) and regarded as a “two-way” fingerprint.

Three-Way Fingerprint PARAFAC Model. An EEFM may be considered as a “two-way” (2D) fingerprint since there are two independent variables, excitation and emission wavelengths, and one dependent variable, the spectral intensity.⁵ A group of EEFMs, which were obtained from P samples, can be stacked into a “three-way” data set, O (P × N × M), which is referred to as a “three-way” fingerprint, and the PARAFAC method can be applied to this multi-way data for exploratory analysis. This method is a generalization of the PCA model for higher order arrays and extracts a number of factors, which are embedded in the data, i.e., in this study, the EEFMs of each sample.

Suppose O (P × N × M) is a three-dimensional data matrix containing the EEFM measurements, where P is the number of samples, and N and M indicate the number of emission and excitation wavelengths, respectively. This matrix may be decomposed by PARAFAC as described in Eq. 3:

\begin{array}{l} O_{p n m} = & \sum_{f = 1}^{F} a_{p f} b_{n f} c_{m f} + e_{p n m} \\ (p = 1, 2, ..., P; n = 1, 2, ..., N; m = 1, 2, ... M) \end{array}

(3)

o_pnm are elements of the O (P×N×M) matrix; F is number of rows in the loadings matrix, representing the number of components used in the PARAFAC model; A (P×F), B (N× F), and C (M×F) with elements a_pf, b_nf, and c_mf are the three loading matrices; e_pnm is the residual term. This trilinear model minimizes the sum of squares of the residuals (e_pnm). The most useful advantage of the PARAFAC model is that its solutions are unique, and for a particular case, the loadings of each factor represent the contribution of a pure component to the complex fluorescence spectrum. Since more than one factor is usually extracted by a PARAFAC model, the sample classification criteria in this study were (i) decompose the data matrix according to the PARAFAC model, and (ii) build a model involving the loadings (sample mode) of the PARAFAC model and the categories of the samples, i.e., the pure and adulterated samples of the RCD and RC TCMs. Several tutorials for the application of PARAFAC to EEFM are available in the literature.^30,31 PARAFAC was modeled with the use of the N-way Tool-box.³²

Classification Models. Four well-known chemometrics methods, KNN, LDA, BP-ANN, and RBF-ANN, were applied for the classification of the different RCD and RC samples and their performance was compared. These methods are well described in the literature,^33–37 but a synopsis of each is provided below.

K-Nearest Neighbors . KNN is a non-parametric method first introduced by Fix and Hodges.³⁸ In this method, a distance (Euclidean or other) is assigned between all points in the data set. The data points, K-closest neighbors (K being the number of neighbors), are then found by analyzing (sorting) the distance matrix. The data set is then sorted into groups of K-closest neighbors. In each class, every data point is allocated a class label. The most common label is then assigned to each data point. In general, it has been suggested that the optimal value of K should be close to √N.^39,40 KNN has several advantages: it is mathematically simple, free from statistical assumptions, and its effectiveness does not depend on the space distribution of the classes.

Linear Discriminant Analysis . LDA is a common method for data classification that finds a linear combination of features best separating two or more classes of object or event. This method maximizes the ratio of between-class variance to the within-class variance in any particular data set. This produces a maximum separation of classes, which is sensitive to the ratio of the number of samples to the number of variables; the ratio is usually between 3∼5.⁴¹ Details about the LDA model can be found in Ref. 40.

Radial Basis Function and Back-Propagation Artificial Neural Networks. In general, an ANN model provides an iterative learning method for the analysis of experimental data.^42,43 It is a powerful chemometrics approach because it does not need any model structure specification and can process multivariate problems of nonlinear systems. With appropriate training, ANNs can accurately model the presence of synergistic effects and avoid the influence of potential nonlinearity in mixtures, which result from the interferences between the components. There are two common ANN methods, which have different transfer functions, i.e., the RBF-ANN and BP-ANN models.⁴⁴ The kernel or basis function is classified as a local activation function compared with the sigmoid function. In this case, the main difference is that with the BP-ANN model the basis function defines an ellipsoid in the input space while with the RBF-ANN a Gaussian function is utilized; it is characterized by two parameters: the center and the peak width (c_j and σ _j ). The output from the jth Gaussian neuron for an input object, x _i , can be calculated by the expression o_j(x) = exp (–|| x _j – c_j||²/σ _j ), where || x _j – c_j|| is the Euclidean distance between x _i and center (c_j).

The output, which computes the weighted sum of the hidden node outputs, is given by

y_{i} = \sum_{i = 1}^{n} w_{j i} o_{j} (x)

(4)

where w_ji represents the weights of the connections between the hidden layer, i, and output layer, j, and the o_j(x) can be obtained from the previously given expression above.

K-nearest neighbors and LDA procedures were carried out with the use of the ChemoAC toolboxes (version 1.1).⁴⁵ The BP-ANN and RBF-ANN algorithms were written with the MATLAB 6.5 software and the subroutines in the Nnet toolbox for MATLAB.

In this investigation, the measured fluorescence data from the RCD-A, RC-A, and AD samples were partitioned into training and validation data sets. In general, two-thirds of the authentic and AD samples were randomly selected for the training data set, i.e., 40 RCD-A, 20 RC-A, and 10 AD samples, and the remaining one-third were included in the validation data set, i.e., 20 RCD-A, 10 RC-A, and 5 AD samples. The quality of the 14 commercial samples, i.e., 11 RCD-C, 3 RC-C samples, was determined based on the resulting discrimination model. Also, importantly the classification performance of the several above-mentioned chemometrics methods was compared.

RESULTS AND DISCUSSION

Spectra and Preprocessing. Under the experimental conditions described in the Instrumentation section, the second-order Rayleigh scattering was eliminated. However, the first-order Rayleigh scattering signals were detected in the measured spectrum. It is generally difficult to minimize this scattering with the use of conventional signal decomposition and/or wave-filtering methods. On the other hand, as the positioning of the Raleigh scattering peaks have been identified,⁴⁶ any corresponding bands found in this study, i.e., any emission bands at wavelengths ≤ the excitation wavelength+20 nm, could be set to zero, and thus, the scatter effects would be minimized. This was adopted as the pretreatment procedure, which removed major background interferences from the EEFM profiles, and enabled access to the spectral information from the TCMs. Such 3D fluorescence spectra and their associated contour maps (Fig. 2; λ_ex = 215– 395 nm and λ_em = 290–560 nm) for the three types of investigated sample, RCD-A (A), RC-A (B), and AD-3 with the RCD : RC of 1 : 1 (C), were clearly different; this is especially evident for RCD and RC. The RCD-A spectra have their two most intense fluorescence peaks (labels 1 and 2) between λ_em = 300–350 nm, whereas there are two less intense peaks (labels 3 and 4) and one strong peak (label 5) between λ_em = 440–500 nm (Fig. 2A). The RC-A spectra show only two fluorescence peaks (labels 6 and 7) between λ_em = 300–400 nm (Fig. 2B); these are somewhat similar to the RCD-A ones in this range but do show some differences in the fine structure. This can be readily observed on the contour maps and will be further discussed in the PARAFAC section. In addition, it can be seen (Fig. 2C) that for the adulterated RC spectra a decrease of fluorescence occurs mainly in the spectral range λ_em = 440– 500 nm compared with what is observed with the RCD spectra; in this case, for the spectral range, λ_em = 300–400 nm, the contour and peak intensity is similar to that found with the RC spectra. However, it is not always easy to detect the spectra of the adulterated RC samples by subjective visual inspection (Fig. 1A), and hence, chemometrics analysis was applied; in general, this approach adds quantitative rigor to data interpretation.

Fig. 2.

3D fluorescence spectra and contour EEFM plots of RCD-A (A), RC-A (B), and AD with RCD and RC in a ratio of 1 : 1 (C). Major fluorescence peaks were labeled 1–7, and spectra were corrected for Rayleigh scattering prior to processing.

Two-Way Spectral Fingerprints of the TCMs Based on Their EEFM. As described in the section Development of Two-way Fingerprints Based on EEFM Data Using PCA, the t₁ element of the score matrix, T (Eq. 2) can be regarded as a two-way spectral fingerprint of the sample. Firstly, PC1 vectors t₁ (score) and v₁ (loadings) are obtained on the basis of EEFM with the use of PCA. In general, each PC1 explained more than about 70% data variance of a data matrix, which indicated that most information in the EEFM data has been successfully extracted. Examples of PC1 plots for scores, t₁, and loadings, v₁, from two randomly selected RCD-A and RC-A samples are shown in Fig. 3; they represent a linear combination of an emission spectrum at each excitation wavelength (t₁), and the weights of each emission spectrum (v₁), respectively. Thus, the plots of the first score vector, t₁, of RCD-A and RC-A correspond to their respective emission spectra, while the plots of the loading vector, v₁, correspond to their respective excitation spectra. This indicates that the maximum excitation wavelengths (at about 220 and 280 nm) produce the largest contribution to t₁. In a sense, the score vector, t₁, can extract the most information from both the excitation and emission spectra; therefore, this vector can be used as a characteristic spectral fingerprint of the sample. Consequently, a group of scores, t₁, each from a different sample, can be combined into a two-way data matrix (O^new(119 × 541)).

Fig. 3.

(A) The first principal component (t₁), and (B) the first loadings vector (v₁) of RCD-A and RC-A samples (solid line: RCD-A; dashed line: RC-A).

Principal Component Analysis on the Two-Way Fingerprint. In order to display a distribution of the 119 samples, PCA was applied to the two-way fingerprint data matrix O^new(119 × 541). This matrix was normalized prior to submission to PCA, with the use of the relationship x' _ij = [x_ij – min(x_j)] / [max(x_j – min(x_j)], where max(x_j) and min(x_j) are respectively the minimum and maximum values of vector x _j . The PC1 versus PC2 projection plot of the two-way fingerprints from the 119 samples (Fig. 1B) indicated that when the objects were projected onto PC1 or PC2, they overlapped significantly in each case and no grouping or clustering could be discerned. However, on the PC1 versus PC2 bi-plot the RCD-A and RC-A objects displayed some separation. In addition, the RC-C commercial samples overlapped the RC-A cluster, which indicated that these samples are similar to the authentic ones. Most of the RCD-C objects are included in the authentic group and any exceptions are identified (circled). These two supposedly commercial RCD samples are located roughly between the RCD and the RC-A groups in the middle of the scattered AD (mixed RCD/RC samples) samples, and hence, the two atypical RCD-C may be adulterated. Thus, PCA has indicated that the two-way fingerprints derived from the EEFM data of the authentic RCD and RC may be discriminated from the commercial samples to some extent. This is a significant improvement when compared to attempts to discriminate the two above analytes visually (Fig. 1A). However, the RCD-A and AD samples also overlap somewhat, and hence, it is still difficult to effectively detect all adulterated samples.

Three-Way Fingerprints and Parallel Factor Analysis. In the Three-Way Fingerprint and PARAFAC section, it was demonstrated how the EEFM spectra may be presented as a three-way fingerprint array for P samples (samples × λ_em × λ_ex). Consequently, the PARAFAC method was applied to resolve the EEFM fingerprints of the three groups: RCD-A (60 samples) and RC-A (30 samples) as well as all of the 119 samples; the dimensions of the three-way arrays were: 60 × 541 × 37 (RCD-A), 30× 541 × 37 (RC-A), and 119× 541× 37 (the whole sample set), respectively. Non-negativity constraints were applied to fit all the PARAFAC models. A CORCONDIA analysis (core consistency diagnostic)⁴⁷ was performed on each PARAFAC model of the above data in order to evaluate the number of significant factors. Three factors were extracted for both the RCD-A and RC-A sample sets, and four for the whole sample set. All fluorescence loadings, including B (emission) and C (excitation) obtained by PARAFAC were normalized with the use of the expression x' _ij = [x_ij – min(x_j)] / [max(x_j – min(x_j)] as previously described above.

The PARAFAC results indicate that there are three fluorescent components for both the RCD and RC sample sets. Their excitation and emission loadings (a1, a2, and a3) are displayed for the RCD samples in Figs. 4A1 and 4A2; for the RC samples, the excitation and emission loadings (b1, b2, and b3) are displayed in Figs. 4B1 and 4B2; and for the whole sample set, the excitation and emission loadings (c1, c2, c3, and c4) are displayed in Figs. 4C1 and 4C2, respectively.

Fig. 4.

PARAFAC excitation and emission loadings of (A) RCD-A, (B) RC-A, and (C) all samples; 99.0%, 90.2%, and 83.6% data variance explained, respectively. Dashed line: the excitation and emission spectra of protopine.

Qualitative comparison of the excitation an emission loadings (Fig. 4) indicates that the loadings profiles a1, b2, and c1 are quite similar to the EEFMs of pure protopine (see Instrumentation section). This suggests that the intense peak in the λ_em = 300–350 nm range observed in the fluorescence spectrum (labels 1 and 2, Fig. 2) can be attributed to the fluorescence of protopine. This assertion is supported by the molecular structure of protopine, which has properties that facilitate fluorescence: it is a polycyclic aromatic hydrocarbon compound with π-conjugation and a rigid structural plane (Scheme 1). Also, it is one of the most effective components with high pharmacological activity and is usually chosen as a marker compound to assess the quality of RCD and RC. In contrast, another marker compound, tetrahydropalmatine, does not have a molecular structure that facilitates fluorescence, and consequently, its spectrum and the related loadings were not observed (Scheme 1B). In summary, PARAFAC modeling of the EEFMs led to the following conclusions: (i) three fluorescent spectral components were resolved for RCD and RC: a1, a2, and a3 for RCD and b1, b2, and b3 for RC; (ii) components a1 and b2 have the same profile as a spectrum of protopine, while factors a3 and b3 are arguably similar to each other but their source(s) are unclear; (iii) factor a2 is specific to RCD (labels 3, 4, and 5, Fig. 2), and factor b1 is specific to RC (labels 6 and 7, Fig. 2); (iv) four, rather then three, factors were extracted for the whole sample set; c1 is similar to a1 and b2 (protopine), c2 has some features similar to a3 and b3, c3 is similar to a2, and c4 is similar to b1 (Fig. 4C).

The loadings matrix, A (119 × 4), which contained the concentrations of the four fluorescent species and was extracted by the PARAFAC model, was plotted vector by vector (i.e., component by component). All plots showed a degree of data discrimination but arguably the best separation of the different classes was displayed on the Component 1 versus Component 4 plot (Fig. 5). However, this plot is similar to the PCA plot (Fig. 1B), which clearly separated the RCD-A and RC-A groups, with the AD sample objects widely scattered between them. This latter group was composed of three different types of sample (see the Materials and Reagents section) with the RCD/RC ratios of 1:4, 4:1, and 1:1. On the three-way fingerprints component-by-component plot (PARAFAC, Fig. 5), these objects seem to form small, discriminated subsets. Also, the five RCD-C samples (circled) were near the authentic RCD-A group, suggesting a similarity of the two types of sample, but importantly, these two groups were separated by two sub-sets of the AD group, presumably from the 4:1 (high Component 1 scores) and the 1:1 (lower Component 1 scores). This suggests that the RCD-C samples were adulterated with RC. The commercial RC samples were still included in the RCA cluster as they were on the PCA plot generated from the two-way fingerprints. In addition, RCD-A objects tend to have the highest scores on Component 1, which suggests that the protopine content in RCD is higher than that in RC. This observation may be quite useful as a control measure for the RCD and RC samples. On the whole, it would appear that the component plots from the data obtained from the two- and three-way fingerprints were quite similar, although the application of the PARAFAC model did seem to produce a more refined data display, which not only enabled the identification of the RCD and RC objects, but also distinguished the adulterated samples from authentic ones. Also, the commercial RCD-C samples were roughly distributed close to the adulterant group, AD, which itself was close to the RCD-C group, and all three overlapped on PC1 to some extent. This indicates that all three types of sample are related to some extent; it is reasonable to suggest that the RCD-C objects, which are separated from the authentic RCD-A by the adulterated AD ones, may be more closely related to the RCD-C objects than to the RCD-A ones; otherwise, the RCD-C samples might have been expected to be closer to the RCD-A ones.

Fig. 5.

Component 1 versus Component 4 bi-plot of the PARAFAC concentration loadings vectors for the 119 samples.

Classification Models. The performance of the KNN, LDA, BP-ANN, and RBF-ANN models was compared for data classification and discussed on the basis of the two-way and three-way fingerprints.

K-Nearest Neighbors and Linear Discriminant Analysis Classification . In general, KNN and LDA classification models are applied for low dimensional data sets; for the LDA method one of the requirements is that the number of input variables must be smaller than the total number of objects. Thus, it is necessary to reduce the number of variables of the data matrix. This can be achieved in a number of ways before modeling, e.g., with the use of PCA, Fourier transform, or feature selection methods. When the two-way fingerprints with 541 variables (O (119×541)) were used as input data, a univariate feature selection method, the Fisher criterion,⁴⁸ was applied. With this method, the variables with highest discriminating properties are expected to be reflected in the highest Fisher criteria, and ten such variables were selected over the spectral range as input for the KNN and LDA models. The three-way fingerprint data, O (119 × 541 × 37), were first treated by PARAFAC, and the loadings matrix, A (119×4), was used for the KNN and LDA modeling; thus, the four factors extracted by PARAFAC enabled each sample to be described by these vectors in four-dimensional space.

The classification of the RCD-A and RC-A objects in the training and validation sets with the use of the KNN and LDA models was satisfactory (classification rates: 93∼100%) but the discrimination performance for the AD samples was poor (classification rates: 0∼50% for training set and 20∼80% for validation set).

Artificial Neural Network Models . For the ANN models, the PCA scores, which represented the two-way fingerprint information, and the PARAFAC scores, which represented the three-way fingerprints, were used as input data, and the classification performance was compared on that basis. Corresponding to the three data groups, RCD-A, RC-A, and AD, the output was three dimensional, and each output was assigned a group label, [1 0 0], [0 1 0], or [0 0 1], respectively. The parameters for the BP-ANN model were set to 10⁻⁸, 1000, 0.3, 0.95, and 15 for goal sum-squared error, epochs, learning rate factor, momentum, and the number of neurons in the hidden layer, respectively. The parameters for the RBF-ANN model were set to 10⁻⁸, 1000, 10, and 60 for goal sum-squared error, epochs, spread radial basis function and radial neurons, respectively.

In general, the classification performance of both models, BP-ANN and RBF-ANN, for the two data sets was satisfactory and recorded better results for the AD samples, both with the training (classification rates: 80∼100%) and validation sets (70∼100%); however, with the two-way fingerprints, the matching classification was as low as 70 to 80%. On the other hand, the three-way data set when modeled by the PARAFAC method produced only one moderately low matching result of 90%, with the others rating no less than 97%.

Quality assessment on the commercial samples based on both the BP-ANN and RBF-ANN models showed that all of the RC-C samples were predicted to correspond to the authentic RC type, and the 11 RCD-C samples were classified into two groups: six authentic RCD and five AD ones as expected. In agreement with the previous results from the other models, this confirmed that these five RCD-C samples were adulterated and also indicated that the three-way fingerprints processed by the ANN models were particularly well suited for the classification of the quality of such complex substances as RCD and RC.

CONCLUSION

An analytical method based on the measurements of excitation-emission fluorescence matrices of examples of real-world complex substances, Rhizoma corydalis decumbentis (RCD) and Rhizoma corydalis (RC), was researched and developed. These matrices were formatted as two- and three-way fingerprints. Qualitative bi-plot analysis with the use of PCA and PARAFAC methods indicated that the authentic RCD-A and RC-A object groups were well separated but the discrimination of the AD from the RCD-C and RC-C samples was less evident, and the PARAFAC modeling produced better results than PCA. Furthermore, loadings plots revealed the presence of the marker compound protopine, a well-known, strongly fluorescing compound, as well as at least two other unidentified significant fluorescent components. This demonstrated that the PARAFAC method can produce analytical evidence that not only supports the established information, i.e., the presence of protopine, but also new information about fluorescent substances that are apparently present but still awaiting identification.

Classification performance of the commonly used methods KNN and LDA was compared with that of the BP-ANN and RBF-ANN models on the basis of two- and three-way formatted data. The common methods performed relatively poorly as compared to the ANN ones, and the best results were obtained with the use of the three-way fingerprints and the RBF-ANN model. Almost 100% classification rates were recorded for RCD-A, RC-A, and AD samples in the training and validation data sets derived from the three-way fingerprints. In addition, all the commercial (RC-C) samples were found to be unadulterated and five adulterated samples were distinguished from the commercial RCD-C ones.

Thus, it has been demonstrated that EEFM spectroscopy is a powerful potential analytical technique for classification analysis of complex real-world substances provided the classification is performed by the RBF-ANN or similar method; this approach is particularly useful where standards for the pure components in complex substances, such as herb reference materials, are unavailable or have not been defined.

Footnotes

ACKNOWLEDGMENTS

The authors gratefully acknowledge the financial support of this study by the National Natural Science Foundation of China (NSFC-21065007), and the State Key Laboratory of Food Science and Technology of Nanchang University (SKLF-MB-201002 and SKLF-TS-200919).

References

J.F.

Fuller

Cattle

Way

C.P.

Hibbert

D.B.

. “Matching Fluorescence Spectra of Oil Spills with Spectra from Suspect Sources”. Anal. Chim. Acta. 2004. 514(1): 51–56.

JiJi

R.D.

Cooper

G.A.

Booksh

K.S.

. “Excitation-Emission Matrix Fluorescence Based Determination of Carbamate Pesticides and Polycyclic Aromatic Hydrocarbons”. Anal. Chim. Acta. 1999. 397(1–3): 61–72.

Arancibia

J.A.

Boschetti

C.E.

Olivieri

A.C.

Escandar

G.M.

. “Screening of Oil Samples on the Basis of Excitation-Emission Room-Temperature Phosphorescence Data and Multiway Chemometric Techniques. Introducing the Second-Order Advantage in a Classification Study”. Anal. Chem. 2008. 80(8): 2789–2798.

Guimet

Ferre

Boque

. “Rapid Detection of Olive–Pomace Oil Adulteration in Extra Virgin Olive Oils from the Protected Denomination of Origin ‘Siurana’ Using Excitation–Emission Fluorescence Spectroscopy and Three-Way Methods of Analysis”. Anal. Chim. Acta. 2005. 544(1–2): 143–152.

Hall

G.J.

Kenny

J.E.

. “Estuarine Water Classification using EEM spectroscopy and PARAFAC–SIMCA”. Anal. Chim. Acta. 2007. 581(1): 118–124.

Díez

Sarabia

Ortiz

M.C.

. “Rapid Determination of Sulfonamides in Milk Samples Using Fluorescence Spectroscopy and Class Modeling with N-Way Partial Least Squares”. Anal. Chim. Acta. 2007. 585(2): 350–360.

Subhash

Wenzel

Lichtenthaler

H.K.

. “Changes in Blue-Green and Chlorophyll Fluorescence Emission and Fluorescence Ratios During Senescence of Tobacco Plants”. Remote Sens. Environ. 1999. 69(3): 215–223.

Nakamura

Tanaka

Ogata

Naruse

. “Dye Analysis of Shosoin Textiles Using Excitation-Emission Matrix Fluorescence and Ultraviolet-Visible Reflectance Spectroscopic Techniques”. Anal. Chem. 2009. 81(14): 5691–5698.

Lakowicz

J.R.

. “Principles of Fluorescence Spectroscopy”. New York. Springer, 2006.

10.

Shi

X.L.

Wei

Y.J.

Zhang

Y.H.

Wang

Y.M.

Liu

C.G.

. “Fluorescence Spectra of Bletilla Striata Aqueous Extraction”. Spectrochim. Acta, Part A. 2007. 27(4): 769–772.

11.

Zhang

Y.H.

Shi

X.L.

Liu

Y.L.

Liu

C.G.

Wei

Y.J.

. “Fluorescence Spectra of Dichroa Febri Fuga Aqueous Extraction”. Spectrochim. Acta, Part A. 2007. 27(2): 342–345.

12.

Ortiz

M.C.

Sarabia

L.A.

Sanchez

M.S.

Gimenez

. “Identification and Quantification of Ciprofloxacin in Urine through Excitation-Emssion Fluorescence and Three-Way PARAFAC Calibration”. Anal. Chim. Acta. 2009. 642(1–2): 193–205.

13.

Maggio

R.M.

Damiani

P.C.

Olivieri

A.C.

. “Four-Way Kinetic-Excitation-Emission Fluorescence Data Processed by Multi-Way Algorithms of Carbaryl and 1-Naphthol in Water Samples in the Presence of Fluorescent Interferents”. Anal. Chim. Acta. 2010. 677(2): 97–107.

14.

Gomez

Callao

M.P.

. “Analytical Applications of Second-Order Calibration Methods”. Anal. Chim. Acta. 2008. 627(2): 169–183.

15.

Escandar

G.M.

Olivieri

A.C.

Faber

N.K.M.

Goicoechea

H.C.

de la Pena

A.M.

Poppi

R.J.

. “Second- and Third-Order Multivariate Calibration: Data, Algorithms and Applications”. TrAC Trends Anal. Chem. 2007. 26(7): 752–765.

16.

H.D.

Wang

Y.J.

Guo

Z.G.

Chang

X.Y.

X.H.

. “Simultaneous Determination of Tetrahydropalmatine, Protopine, and Palmatine in Rat Plasma by LC-ESI-MS and Its Application to a Pharmacokinetic Study”. J. Pharm. Biomed. Anal. 2009. 49(2): 440–446.

17.

Sheng

Z.L.

Jiang

Cao

. “Effects of Rhizoma Corydalis Decumbentis on Learning and Memory Ability and Brain Acetylcholinesterase Activity in Mice”. Chin. Tradit. Herb. Drug. 2003. 34: 543–545 (in Chinese).

18.

Janbaz

K.H.

Saeed

S.A.

Gilani

A.H.

. “An Assessment of the Potential of Protopine to Inhibit the Microsomal Drug Metabolizing Enzymes and Prevent Chemical-Induced Hepatotoxicity in Rodents”. Pharmacol. Res. 1998. 38(3): 215–219.

19.

Cheng

Z.H.

Guo

Y.L.

Wang

H.Y.

Chen

G.Q.

. “Qualitative and Quantitative Analysis of Quaternary Ammonium Alkaloids from Rhizoma Corydalis by Matrix-Assisted Laser Desorption/Ionization Fourier Transform Mass Spectrometry Coupled with a Selective Precipitation Reaction Using Reinecke Salt”. Anal. Chim. Acta. 2006. 555(2): 269–277.

20.

Chinese Pharmacopoeia Commission. “Pharmacopoeia of the People's Republic of China”. Beijing. Chinese Chemical Industry Press, 2005. Pp. 94 and 197.

21.

Saeed

S.A.

Gilani

A.H.

Majoo

R.U.

Shah

B.H.

. “Anti-Thrombotic and Anti-Inflammatory Activities of Protopine”. Pharmacol. Res. 1997. 36(1): 1–7.

22.

F.N.

T.S.

S.T.

Y.C.

Huang

T.F.

Teng

C.M.

. “Antiplatelet Effects of Protopine Isolated from Corydalis Tubers”. Thromb. Res. 1989. 56(2): 289–298.

23.

Leung

W.C.

Zheng

Huen

Law

S.L.

Xue

. “Anxiolytic-Like Action of Orally Administered Dl-Tetrahydropalmatine in Elevated Plus-Maze. Prog. Neuropsychopharmacol”. Biol. Psychiat. 2003. 27(5): 775–779.

24.

Hsu

Kin

K.C.

. “Pharmacological Study of Tetrahydropalmatine and Its Analogs, a New Type of Central Depressants”. Arch. Int. Pharmacodyn. 1962. 139: 318–327.

25.

Jiang

Cao

Wang

. “Inhibitory Effect of Protopine on K_ATP Channel Subunits Expressed in HEK-293 Cells”. Eur. J. Pharmacol. 2004. 506(2): 93–100.

26.

Lin

M.T.

Wang

J.J.

Young

M.S.

. “The Protective Effect of Dl-Tetrahydropalmatine Against the Development of Amygdala Kindling Seizures in Rats”. Neurosci. Lett. 2002. 320(3): 113–116.

27.

Sturm

Seger

Stuppner

. “Analysis of Central European Corydalis Species by Nonaqueous Capillary Electrophoresis–Electrospray Ion Trap Mass Spectrometry”. J. Chromatogr. A. 2007. 1159(1–2): 42–50.

28.

Martens

Naes

. “Multivariate Calibration”. New York: John Wiley and Sons, 1991.

29.

Yan

S.K.

Xin

W.F.

Luo

G.A.

Wang

Y.M.

Cheng

Y.Y.

. “An Approach to Develop Two-Dimensional Fingerprint for the Quality Control of Qingkailing Injection by High-Performance Liquid Chromatography with Diode Array Detection”. J. Chromatogr. A. 2005. 1090(1–2): 90–97.

30.

Bro

. “PARAFAC. Tutorial and Applications”. Chemom. Intell. Lab. Syst. 1997. 38(2): 149–171.

31.

Gimenez

Sarabia

Ortiz

M.C.

. “The Maintenance of a PARAFAC Calibration and the Second-Order Property: Application to the Determination of Ciprofloxacin in Presence of Enrofloxacin by Excitation– Emission Fluorescence”. Anal. Chim. Acta. 2005. 544(1–2): 327–336.

32.

Andersson

Bro

. “The N-Way Toolbox”. http://www.models.kvl.dk/source/nwaytoolbox/.

33.

Derde

M.P.

Buydens

Guns

Massart

D.L.

Hopke

P.K.

. “Comparison of Rule-Building Expert Systems with Pattern Recognition for the Classification of Analytical Data”. Anal. Chem. 1987. 59(14): 1868–1871.

34.

Coomans

Massart

D.L.

. “Alternative K-Nearest Neighbor Rules in Supervised Pattern Recognition. Part 2. Probabilistic Classification on the Basis of the KNN Method Modified for Direct Density Estimation”. Anal. Chim. Acta. 1982. 138(1): 153–165.

35.

Coomans

Jonckheer

Massart

D.L.

Broeckaert

Blockx

. “The Application of Linear Discriminant Analysis in the Diagnosis of Thyroid Disease”. Anal. Chim. Acta. 1978. 103(4): 409–415.

36.

Persaud

K.C.

Byun

H.G.

. “Classification of Complex Odours Using Polymer Sensor Arrays and Neural Networks”. In: Soulie

F.F.

Gallinari

, editors. Industrial Applications of Neural Networks. Singapore: World Scientific, 1998. Pp. 85–90.

37.

Byun

H.G.

Persaud

K.C.

Kim

J.D.

Lee

D.D.

. “Improvement of RBF Network's Performance for Application to Odour Sensing Systems”. In: Weimar

Frank

, editors. Proc. of ISOEN99 (Tubingen, Germany, Sept. 20-22). 1999. Pp. 237–240.

38.

Fix

Hodges

J.L.

. “Discriminatory Analysis - Nonparametric Discrimination: Consistency Properties”. Int. Stat. Rev. 1989. 57(3): 238–247.

39.

Amendolia

S.R.

Cossu

Ganadu

M.L.

Golosio

Masala

G.L.

Mura

G.M.

. “A Comparative Study of K-Nearest Neighbour, Support Vector Machine and Multi-Layer Perceptron for Thalassemia Screening”. Chemom. Intell. Lab. Syst. 2003. 69(1–2): 13–20.

40.

Sanchez

M.S.

Sarabia

L.A.

. “Efficiency of Multi-Layered Feed-Forward Neural Networks on Classification in Relation to Linear Discriminant Analysis, Quadratic Discriminant Analysis and Regularized Discriminant Analysis”. Chemom. Intell. Lab. Syst. 1995. 28(2): 287–303.

41.

Casale

Sinelli

Oliveri

Di Egidio

Lanteri

. “Chemometrical Strategies for Feature Selection and Data Compression Applied to NIR and MIR Spectra of Extra Virgin Olive Oils for Cultivar Identification”. Talanta. 2010. 80(5): 1832–1837.

42.

Looney

C.G.

. “Pattern Recognition Using Neural Networks”. New York: Oxford University Press, 1997. pp. 75.

43.

Zupan

Gasteiger

. “Neural Networks for Chemists: an Introduction”. Weinheim: VCH, Germany, 1993.

44.

Berrueta

L.A.

Alonso-Salces

R.M.

Heberger

. “Supervised Pattern Recognition in Food Analysis”. J. Chromatogr. A. 2007. 1158(1–2): 196–214.

45.

http://www.vub.ac.be/fabi/toolboxes.html.

46.

Jiang

F.H.

Lee

F.S.C.

Wang

X.R.

Dai

D.J.

. “The Application of Excitation/Emission Matrix Spectroscopy Combined with Multivariate Analysis for the Characterization and Source Identification of Dissolved Organic Matter in Seawater of Bohai Sea, China”. Mar. Chem. 2008. 110(1–2): 109–119.

47.

Bro

Kiers

H.A.L.

. “A New Efficient Method for Determining the Number of Components in PARAFAC Models”. J. Chemom. 2003. 17(5): 274–286.

48.

Mallet

Walczak

Penninckx

Massart

D.L.

Heuerding

Erni

. “Comparison of Regularized Discriminant Analysis, Linear Discriminant Analysis and Quadratic Discriminant Analysis”. Anal. Chim. Acta. 1996. 329(3): 257–265.