中文摘要

高效液相層析儀(HPLC)及感應耦合電漿質譜儀（ICP-MS）為測定中藥指標成分及微量元素的常用分析工具。本研究開發分析方法，進行藥材的成分分析及樣品辨識。
本研究共分三部份，第一部份為利用加味逍遙散HPLC的有機成分(30 peaks)定量結果及ICP-MS的無機元素(44個元素)測定數據，以人工歸納及統計分析方法建立識別廠牌的準則。觀察各樣品有機成分中成分2 ( Gallic acid ，GA ) 及成分26 ( Paeonol ，PN )的差異，發現利用成分2/IS比值可明顯將所有樣品分成兩組，順天堂、勝昌、科達其值高於1.45，而港香蘭、莊松榮低於1.45；繼而以2/26的比值(順天堂>2.1、1.8<勝昌<2.1、科達<1.8、港香蘭<2.3、莊松榮>2.5)再行細分，則各家廠牌產品就可順利分辨。依上述準則，在五家廠牌共33批樣品中，只有4批樣品分類錯誤：勝昌2批、港香蘭1批其2/IS比值分別為2.21、2.37和1.67，且2/26的比值高於2.1被判定為順天堂；順天堂1批其2/IS比值為1.65，且2/26的比值在1.8與2.1之間被判定為勝昌，整體而言，人工辨識正確率約88 ％。將30 peaks的分析數據進行多變量分析(Multivariate Analysis)處理：主成分分析(Principal Component Analysis ，PCA)發現，PC1(1、11和14)、PC2(2、6、24和25)和PC3(23、27和28)代表61.343% 的總變異性；以PC1、PC2和PC3作成的三度空間PCA圖顯示，各廠牌樣品各自群聚在特定位置，順天堂在中間偏上方，勝昌在左下方，科達在中間，港香蘭在右下方，莊松榮在中間偏下方，分類良好。進行群聚分析(Cluster Analysis (CA)亦能明確分類，在群聚距離設定為0.38（最大刻度為1）後，依次出現順天堂、莊松榮、港香蘭、勝昌、科達等五聚落。進行線性區別分析(Linear Discriminant Analysis ，LDA)， 其原始分類(original)雖達100％，但在交叉驗證(cross-validated)卻只有13.8 %的正確性，因此先用逐步區別分析(stepwise-LDA)選出1(nicotinic acid)、 8(Geniposide)、13 、17 、19、20 、23 、24等8 peaks，再用LDA統計則可得到完全正確的分類結果，該方法亦可正確判斷任一市售樣品的廠牌，另外，17和19在判定函數1中的變數有最大的比重(-26.620 和 –18.404)，判定函數2主要由20和24 (-11.716和-12.520)所組成，判定函數3主要由17、23和24 (21.550、20.372和-17.723)所組成，判定函數4主要由17、24和20 (-8.869、-6.954和6.051)所組成。
用ICP-MS測量各樣品，可得到44個無機元素的分析數據，仔細比對這些數據的差異性，僅能分辨順天堂、勝昌及港香蘭三家產品，而科達及莊松榮的所有元素含量都非常接近，且誤差線範圍內有很大的重疊，人工難以辨別。將該44無機元素的分析數據進行多變量分析：主成分分析顯示，PC1(La、Ce、Nd和Pr)、PC2(Al、Mg和Br)和PC3(I和As)代表71.944% 的總變異性，以PC1、PC2和PC3繪製的的三度空間PCA圖指出，順天堂在左下方，勝昌在左上方，科達在右下方，港香蘭在右上方，莊松榮在中間偏下方，各自形成聚落、分隔明顯。群聚分析以距離0.41（最大刻度為1）處理後，明確分成五個聚落，依次為順天堂、莊松榮、港香蘭、科達、勝昌。線性區別分析，其原始分類可達100％，但交叉驗證只有17.2 %的正確性，因此先以逐步區別分析自44個元素中選出Na、Mg、Al、Ni、Cu、I、Ce、Bi、U等9個元素，再用LDA統計則可得到完全正確的分類結果，所有的加味逍遙散樣品皆能無誤地分類，該統計分析方法亦可正確判斷任一市售樣品的廠牌，其中 Mg、Na和Cu(-2.443、2.203和2.140)在判定函數1中的變數有最大的比重，判定函數2主要由Ce和U (-6.208和4.730)所組成，判定函數3主要由Cu、Mg和I (1.843、-1.126和1.063)所組成，判定函數4主要由U、Mg和Ce (1.185、1.115和-1.021)所組成。
另外，隨機抽取三個樣品，以比較五家廠牌的同質性，發現順天堂、莊松榮及勝昌生產之不同批次樣品，相似度最高，分別為0.23、0.22及0.21，品質最穩定，港香蘭及科達所生產之商品相似度較差，分別為0.16及0.09，品質較不穩定。
第二部份為枳實(Aurantii Fructus Immaturus)及枳殼(Aurantii Fructus Maturus)之分析與辨識。從台灣及中國的中藥市場收集20批枳實[綠衣枳實(Poncitrus trifoliata)]和30批[酸橙枳殼 (Citrus aurantium)，香圓枳殼(C. wilsonii)]樣品，經外觀判別及組織鏡檢，再以HPLC分析這些樣品中的12個主要的成分；該HPLC分析方法以Cosmosil 5C18-AR為分析管柱，以KH2PO4水溶液及MeOH/CH3CN之混合液為沖提液，可在60分鐘內順利完成分離，積分面積的RSD為0.30-1.21(intraday) 及0.45-1.93 (interday) ，偵測極限在2.84-6.85 ng之間。分析結果顯示，枳實與枳殼可用NR、 HE、 NE和QU等成份之有無來判別，枳實樣品中沒有或幾乎沒有HE及NE等成份，而枳殼樣品中沒有或很少NR及QU成份；兩枳殼(酸橙及香圓)的差異，則可用NGC及HE的成分的含量及HE/NG，HE/NE與NGC/NE比值分辨， 該等比值在酸橙枳殼分別為0.49以上、2.30以上及0.21以上，在香圓枳殼則在0.40以下、1.31以下及0.21以下，但有5批不易判別，依相近數據，判定分屬兩品種。將12 peaks的分析數據代入多變量分析軟體，發現不論PCA、CA或LDA均能明確分辨綠衣枳實、酸橙枳殼或香圓枳殼，經交叉驗證或添加試驗，均得到完全正確的結果。值的注意的是，5批用人工不易判別的樣品皆清楚落入香圓枳殼的群聚，與人工辨識有著很大的不同。
第三部份，我們利用本實驗室已有的HPLC分析方法，測定牡丹皮(Moutan Cortex) 、厚朴(Magnoliae Cortex)、防己(Fangchi Radix)等的主要有機成分及微量元素含量。我們曾比較水煮及消化兩種前處理方式，發現後者約為前者的102到104倍，但所有數值皆呈平行關係，因此選擇與中藥煎煮情形相同的水煮方法，製作檢液。
1.牡丹皮(Moutan Cortex):
收集23批市售牡丹皮樣品，分屬西昌丹皮（Paeonia delavayi，10批）、川丹皮（P. suffruticosa，13批），用HPLC測量4,6-di-GG (1，G為glucose)，1,2,3,6-tetra-GG (2) ，1,2,3,4,6-penta-GG (3) ， 1,3,4,6-tetra-GG (4) ，3,4,6-tri-GG (5) ，1,3,6-tri-GG (6) ， 3,6-di-GG (7) ，1,2,6-tri-GG (8) 。paeoniflorin（Pf）和paeonol(PN)等10成分含量。發現由PN/Pf及3/2的比值就可區別西昌丹皮(PN/Pf > 1， 3/2>15.1)與川丹皮(PN/Pf < 1， 3/2<11.4) ；整體而言，人工辨識正確率為100％。將10 peaks分析數據輸入PCA、CA及LDA軟體 ，發現主成分分析與線性區別分析，都能正確分辨各批藥材所屬的品種，後者的交叉驗證亦證實100 %正確。但群聚分析則無法分辨，因此以逐步區別分析選出peak1(4,6-di-GG)、peak4(1,3,4,6-tetra-GG)、Pf(paeoniflorin))等3成分，再進行群聚分析，但也無法正確分類；不過只用該三peaks的定量數據，進行線性區別分析，卻也可得到100%的正確結果，顯示用LDA統計分析時，不論10成分或3成分都能明確分辨基原，但後者更為簡捷。
仔細觀察18批牡丹皮樣品(西昌丹皮9批、川丹皮9批)的62種元素之分析數據，發現Ge、Mo、Cd、Ce及Tl等五元素，在兩品種中相異性最顯著，用該五元素的分析數據作為人工辨識的工具，正確率達100％。若將62種元素分析數據用主成分分析、群聚分析及線性區別分析軟體處理，則只有PCA可達到完全正確的分類目標，CA無法分辨，LDA之原始分類雖有100%正確性，但交叉驗證卻只有50%的分辨率。我們用stepwise-LDA選出Be、V、Ag、Cd、Tl、Pb、Bi及Th等8元素，用該8元素為變數，則CA及LDA兩統計分析模式，均能完全確認所有樣品的基原。
2.厚朴(Magnoliae Cortex)：
收集22批分屬川厚朴（Magnolia officinalis Rehd. Et Wils，12批）、凹葉厚朴 [M. officinalia ssp. biloba （Rehd. et Wils.）Law，4批]、和厚朴(M. obovata THUNB. ，6批)的市售藥材，用HPLC測量(-)-magnocurarine (1)， (+)-magnoflorine (2)，(+)-laurifoline (3)，(+)-oblongine (4)，(+)-menisperine (5)，(+)-xanthoplanine(6)，(+)-N-methylglaucine (7)等七個生物鹼及Honokiol (8)， Magolol(9)兩個酚類成分含量。發現由2/1及9/8的比值可挑出屬於川厚朴(2/1>5.102及9/8<1.397)的樣品，再由2/3及4/3的比值就可區別凹葉厚朴(2/3<6.458及4/3<1.130)與和厚朴(2/3>18.901及4/3>3.156)，用這流程的人工辨識正確率為100％。將9 peaks的定量結果輸入PCA、CA及LDA軟體，發現各統計方法都能迅速歸納出基原與成分的關係，分辨結果完全正確；以 stepwise-LDA處理，選出7及8兩成分作為新變數，結果顯出只用該二吸收峰的定量數據，亦能精確指示任一樣品所屬的基原。
用收集9批厚朴樣品(川厚朴7批、凹葉厚朴2批)的46個無機元素之分析數據，經仔細比對其差異性，發現Fe、Pb、V、Zr、Sb、Bi、Th等七元素可作為人工分辨依據。我們將46個元素的分析數據，以PCA、 CA及LDA處理，結果顯示PCA及LDA均能正確分類，CA則無法明顯劃分群落；用stepwise-LDA自46個元素中選出Sc、Cu、Sr、Zr和Th等5個元素，以該5數據為變數，不論CA或LDA均能正確分辨，前者至距離0.31(最大刻度為1)就可將川厚朴與凹葉厚朴分成兩群落，後者用五元素與四十六元素均能得到100%的交叉驗証。
3.防己（Fangchi Radix）：
收集37批分屬日防己(Sinomenium acutum 5批)、粉防己（S. tetrandra 11批）、廣防己（Aristolochia fangchi 15批）、木防己(Cocculus trilobus 6批)的防己樣品，用HPLC分析各藥品發現日防己含有 acutumidine (1)、 magnoflorine (2) 、 stepharine (3) 、 sinomenine (4) 、 acutumine (5) and tetrandrine (9)， 粉防己含有 sinomenine (4) 、 cyclanoline (6) 、 fangchinoline (7) 、 berbamine (8) 、 tetrandrine (9) and isotetrandrine (10)， 廣防己含有magnoflorine (2) 、 tetrandrine (9) 、 aristolochic acid II (12) 、 aristolochic acid I (13) 、 aristololactam (14) 和未知成份 X(推測應該是aristoloactone)， 木防己含有magnoflorine (2) 、 sinomenine (4) 、 tetrandrine (9) 、 trilobine (11) 、 isotrilobine (15),和未知成份 A and B ，各不同品種藥材可由指紋圖譜分辨。以18 peaks分析數據進行多變量分析，發現不論主成分分析、群聚分析或線性區別分析，均能將各樣品依所屬基原迅速分類，正確性及相似度都相當明確，也可用於判定未知樣品的基原。以逐步區別分析自18 peaks中選出2、4、7、11、15等5成分，再用LDA統計亦可得到完全正確的分類結果。
用ICP-MS測量16批防己樣品(日防己2批、粉防己4批、廣防己10批)的55微量元素含量，發現其中Be、Mn、Br、Yb、Tl、U、Bi的差異性，可作為人工分辨依據。將全部55元素數據輸入多變量分析軟體，發現只有主成分分析能達到分辨基原的目標，線性區別分析原始分類雖為100%正確，但交叉驗證只有12.5%的分類結果。我們繼而用逐步區別分析(stepwise-LDA)挑選出Be、Mn、Br、Yb、Tl、U、Bi等7元素作為變數，再進行群聚分析及線性區別分析，結果顯示兩種方法都可以得到完全正確的分類效果。

ABSTRACT

High-performance liquid chromatography (HPLC) and inductively coupled plasma - mass spectrometry (ICP-MS) are analysis tools frequently used to determine the marker substances and trace elements of Chinese herbal medicine. This study has developed methods to carry out the ingredient analysis and sample recognition of medicinal materials.
This research consists of three parts: the first part, through the HPLC quantitative results (30 peaks) of organic ingredients and the ICP-MS assay data of inorganic elements (44 elements) of Chia-wei-hsiao-yao-san (Bupleurum and Peony Formula), adopted manual induction and statistical analysis method to set up rules for distinguishing the brands. The differences of component 2 ( Gallic acid, GA ) and component 26 ( Paeonol, PN ) of the organic ingredients of each samples were observed and it was found that all of the samples could be obviously distinguished into two groups by utilizing the ratio of component 2/IS. The values of Sun Ten Pharmaceutical, Sheng Chang Pharmaceutical and KODA Pharmaceutical were greater than 1.45 while Kaiser Pharmaceutical and Chuang Song Zong Pharmaceutical were lower than 1.45. If they were subdivided by the ratio of 2/26 (Sun Ten > 2.1, 1.8 < Sheng Chang <2.1, KODA <1.8, Kaiser <2.3, Chuang Song Zong > 2.5), the products of each brand would then be recognized smoothly. According to the above rules, only 4 batches out of 33 samples from five brands were classified incorrectly: 2 batches of Sheng Chang and 1 batch of Kaiser with the 2/IS ratio at 2.21, 2.37 and 1.67 respectively and the 2/26 ratio greater than 2.1 was determined as Sun Ten. One batch of Sun Ten with the 2/IS ratio at 1.65 and the 2/26 ratio between 1.8 and 2.1 was determined as Sheng Chang. Overall, the accuracy of recognition by hand was about 88%. The Multivariate Analysis was performed on the analytical data of 30 peaks: it was found in the principal component analysis (PCA) that PC1 (1, 11 and 14), PC2 (2, 6, 24 and 25) and PC3 (23, 27 and 28) represented 61.343% of total variability. The three-dimensional PCA chart formed by PC1, PC2 and PC3 showed that the samples of each brand clustered at specific position respectively: Sun Ten was in the center top position, Sheng Chang in the lower left corner, KODA in the center, Kaiser in the lower right corner and Chuang Song Zong in the center bottom with good classification. They also could be clearly categorized by carrying out the cluster analysis (CA). After the cluster distance was set at 0.38 (the maximum scale is 1), five clusters of Sun Ten, Chuang Song Zong, Kaiser, Sheng Chang and KODA appeared successively. While the linear discriminant analysis (LDA) was performed, the original classification reached 100%. However, the accuracy only resulted in 13.8% when it was cross-validated. Therefore, stepwise-LDA was firstly adopted to select 8 peaks of 1(nicotinic acid), 8(Geniposide), 13, 17, 19, 20, 23 and 24. Then the LDA was used to compile the statistics and the completely correct results of classification could be obtained. This method could also be used to determine accurately the brand of any sample in the market. In addition, the variables 17 and 19 possessed the highest weights in discriminant function 1 (-26.620 and –18.404). The discriminant function 2 was constituted mainly by 20 (-11.716) and 24 (-12.520). Discriminant function 3 was constituted mainly by 17 (21.550), 23 (20.372), and 24 (–17.723) and discriminant function 4 was constituted mainly by 17 (-8.869), 24 (-6.954) and 20 (6.051).
While the ICP-MS was used to test each sample, the analysis data of 44 inorganic elements could be obtained. The differences of these data were compared carefully and only the products of three brands could be distinguished: Sun Ten, Sheng Chang and Kaiser. The contents of all elements of KODA were extremely close to the ones of Chuang Song Zong and the large overlapping within the tolerance was very difficult to be distinguished by hand. The multivariate analysis was conducted on the analysis data of these 44 inorganic elements: the PCA showed that PC1 (La, Ce, Nd and Pr), PC2 (Al, Mg and Br) and PC3 (I and As) represented 71.944% of total variability. The three-dimensional PCA chart formed by PC1, PC2 and PC3 showed that Sun Ten was in the lower left corner, Sheng Chang in the upper left corner, KODA in the lower right corner, Kaiser in the upper right corner and Chuang Song Zong in the center bottom with separate clusters in obvious partition. After the distance of cluster analysis was set at 0.41 (the maximum scale is 1), they were clearly divided into five clusters: Sun Ten, Chuang Song Zong, Kaiser, KODA and Sheng Chang successively. While the linear discriminant analysis was performed, the original classification could reach 100%. However, the accuracy only resulted in 17.2% when it was cross-validated. Therefore, the stepwise-LDA was firstly adopted to select 9 elements from these 44 elements: Na, Mg, Al, Ni, Cu, I, Ce, Bi and U. Then the LDA was used to compile the statistics and the completely correct results of classification could be obtained. All of the Chia-wei-hsiao-yao-san samples could be properly classified and this statistical analysis method could also be used to determine correctly the brand of any sample in the market. In addition, the variables of Mg, Na and Cu (-2.443, 2.203 and 2.140) possessed the highest weights in discriminant function 1. The discriminant function 2 was constituted mainly by Ce (-6.208) and U (4.730). Discriminant function 3 was constituted mainly by Cu (1.843), Mg (-1.126) and I (1.063), and discriminant function 4 was constituted mainly by U (1.185), Mg (1.115) and Ce (-1.021).
Furthermore, three samples were randomly selected to compare the product consistency of these five brands and it was found that the samples from different batches produced by Sun Ten possessed the highest conformity which resulted in 0.23 with the most stable quality. The ones of Chuang Song Zong and Sheng Chang were in the second place which resulted in 0.22 and 0.21 respectively. The commodities produced by Kaiser and KODA had poorer conformity which represented 0.16 and 0.09 respectively with the quality rather unstable.
The second part comprised the analysis and discrimination of Aurantii Fructus Immaturus and Aurantii Fructus Maturus. The samples including 20 batches of Immaturus (Poncitrus trifoliata) and 30 batches of Maturus (Citrus aurantium and C. wilsonii) were collected from the herbal medicine markets in Taiwan and China. After
identifying by the external appearance and pharmacognostic histological anatomy, the HPLC was performed to analyze the 12 principle components in these samples. The Cosmosil 5C18-AR was used as the analysis column and the mixture of aq. KH2PO4 and MeOH/CH3CN was used as elution in the HPLC analytical method. The separation could be achieved smoothly within 60 minutes. The RSD of peak area were 0.30-1.21 (intraday) and 0.45-1.93 (interday). The detecting limits were between 2.84 and 6.85 ng. The analysis results showed that immature samples and mature samples could be distinguished by the constituents of NR, HE, NE and QU. The immature samples did not or hardly contain HE or NE while the mature ones did not contain or contained very few constituents of NR or QU. The difference between both mature samples (C. aurantium and C. wilsonii) could be distinguished by the constituents of NGC and HE as well as the ratio of HE/NG, HE/NE and NGC/NE. The ratios in C. aurantium were above 0.49, 2.30 and 0.21 while the ones in C. wilsonii were 0.40, 1.31 and below 0.21 respectively. However, 5 batches which were not easy to be distinguished and were recognized as two species according to the close data. The analytical data of 12 peaks were entered into the multivariate analysis software and it was found that whether the PCA, CA or LDA could definitely distinguish the P. trifoliate, C. aurantium and C. wilsonii. The completely correct results could be achieved in the cross validation or adjunction experiment. It was notable that the 5 batches of samples not easy to be distinguished by hand were clearly grouped into the cluster of C. wilsonii.
In the third part, we used the HPLC method available in our laboratory to measure the contents of major constituents and trace elements of Moutan Cortex, Magnoliae Cortex and Fangchi Radix. We had compared both pre-treatments of boiling and digestion and found that the latter was around 102 to 104 times of the former. However, all of the values were in parallel correlation. Therefore, we decided to use the boiling method which was equivalent to the decoction of Chinese herbal medicine for sample preparation.
1. Moutan Cortex
The samples of 23 batches of Moutan Cortex sold in the market were collected: 10 batches of Paeonia delavayi and 13 batches of P. suffruticosa. The HPLC was conducted to measure the contents of 10 constituents: 4,6-di-GG(1, G=glucose), 1,2,3,6-tetra-GG(2), 1,2,3,4,6-penta-GG(3), 1,3,4,6-tetra-GG(4), 3,4,6-tri-GG (5), 1,3,6-tri-GG(6), 3,6-di-GG(7), 1,2,6-tri-GG(8), paeoniflorin(Pf) and paeonol(PN). It was found that the ratios of PN/Pf and 3/2 could be adopted to distinguish P. delavayi (PN/Pf > 1, 3/2>15.1) and P. suffruticosa (PN/Pf < 1, 3/2<11.4). Overall, the accuracy of recognition by hand was 100%. The analytical data of 10 peaks were entered into the software of PCA, CA and LDA, and it was found that the species of each batch of drug materials could be correctly distinguished in both principle component analysis and linear discriminant analysis. The cross validation of the latter also proved 100% accuracy, but they could not be discriminated in cluster analysis. Therefore, the stepwise-LDA was performed to select three constituents: peak1(4,6-di-GG), peak4(1,3,4,6-tetra-GG) and Pf (paeoniflorin) and then the cluster analysis was carried out. However, the correct classification could not be obtained either. Nevertheless, performing the linear discriminant analysis by using the analysis data of only three peaks could achieve 100% correct results. This showed that the statistical analysis of LDA could clearly distinguish the origins regardless of 10 or 3 constituents, but the latter is simple and speedy.
The analytical data of 62 elements of 18 batches of Moutan Cortex samples (9 batches of P. delavayi and 9 batches of P. suffruticosa) were observed carefully and it was found that five elements: Ge, Mo, Cd, Ce and Tl possessed the most noticeable heterology in these two species. When the analytical data of these five elements were used as the tool of recognition by hand, the accuracy could reach 100%. If the analytical data of 62 elements were processed by the software of PCA, CA and LDA, only PCA could achieve the goal of completely correct classification and they could not be distinguished in CA. Although the original classification of LDA possessed 100% accuracy, the discrimination ratio only reached 50% in cross validation. We used stepwise-LDA to select 8 elements with the optimal discrimination effects: Be, V, Ag, Cd, Tl, Pb, Bi and Th. Let the 8 elements be variables and both CA and LDA analysis modes could completely confirm the origins of all samples.
2. Magnoliae Cortex
22 batches of Magnoliae Cortex samples were collected: 12 batches of Magnolia officinalis, 4 batches of M. officinalia ssp. biloba（Rehd. et Wils.）and 6 batches of M. obovata THUNB. The HPLC was performed to measure the contents of seven alkaloids: (-)-magnocurarine(1), (+)-magnoflorine(2), (+)-laurifoline(3), (+)-oblongine(4), (+)-menisperine(5), (+)-xanthoplanine(6), (+)-N-methylglaucine(7) and two phenols: honokiol(8) and magolol(9). It was found that the ratios of 2/1 and 9/8could sort out the samples that belonged to M. officinalis (2/1>5.102 and 9/8<1.397) while the ratios of 2/3 and 4/3 could distinguish M. officinalia ssp. Biloba (2/3<6.458 and 4/3<1.130) and M. obovata (2/3>18.901 and 4/3>3.156). The accuracy of recognition by hand adopting this procedure could reach 100%. The quantitative results of 9 peaks were entered into the software of PCA, CA and LDA. It was found that each statistical method could rapidly induce the correlation between the origin and constituent with completely correct discrimination. The stepwise-LDA was performed to select two constituents 7 and 8 as new variables. The result showed that the origin to which any sample belonged could be determined accurately by only adopting the analysis data of two absorption peaks.
The analytical data of 46 inorganic elements of 9 batches of Magnoliae Cortex samples (7 batches of M. officinalis and 2 batches of M. officinalia ssp. biloba) were collected. The differences were compared carefully and it was found that seven elements: Fe, Pb, V, Zr, Sb, Bi and Th could be the basis of recognition by hand. We used PCA, CA and LDA to process the analytical data of these 46 elements, and the results showed that they could be correctly classified by PCA and LDA, but clusters could not be evidently divided by CA. Thus the stepwise-LDA was performed to select 5 from these 46 elements: Sc, Cu, Sr, Zr and Th. These 5 data were taken as variables and no matter CA or LDA could distinguish them correctly. The former could divide M. officinalis and M. officinalia ssp. biloba into two clusters at the distance of 0.31 (the maximum scale is 1) while the latter could obtain 100% cross validation by using both five elements and forty-six elements.
3. Fangchi Radix
37 batches of Fangchi Radix samples were collected: 5 batches of Sinomenium acutum、11 batches of S. tetrandra、15 batches of Aristolochia fangchi and 6 batches of Cocculus trilobus. The HPLC was carried out to analyze each drug material and was found that S. acutum contained acutumidine (1), magnoflorine (2), stepharine (3), sinomenine (4), acutumine (5), and tetrandrine (9); S. tetrandra has sinomenine (4), cyclanoline (6), fangchinoline (7), berbamine (8), tetrandrine (9), and isotetrandrine (10); A. fangchi contains magnoflorine (2), tetrandrine (9), aristolochic acid II (12), aristolochic acid I (13), aristololactam (14), and an unknown component X; and, C. trilobus has magnoflorine (2), sinomenine (4), tetrandrine (9), trilobine (11), isotrilobine (15), and unknown components A and B. Each drug material of different species could be distinguished by above fingerprint. The analytical data of 18 peaks were adopted to perform the multivariate analysis and it was found that all of the PCA, CA and LDA could rapidly classify each sample according to the origin where it belonged with precise accuracy and similarity. The origin of unknown sample could also be determined. The stepwise-LDA was used to select 4 constituents 2, 4, 7, 11 and 15 from 18 peaks. Then the LDA was performed to compile the statistics and the completely correct results of classification could also be obtained.
The ICP-MS was carried out to measure the contents of 55 trace elements of 16 batches of Fangchi Radix samples (2 batches of S. acutum, 4 batches of S. tetrandra and 10 batches of A. fangch). It was found that the differences between Be, Mn, Br, Yb, Tl, U and Bi could be the basis of recognition by hand. When the analytical data of 55 elements were entered into the multivariate analysis software, it was found that only PCA could achieve the goal of origin discrimination. Although the original classification of LDA possessed 100% accuracy, the discrimination result only reached 12.5% in cross validation. Thus we used stepwise-LDA to select 7 elements to be the variables: Be, Mn, Br, Yb, Tl, U and Bi. Then the CA and LDA were performed and the results showed that both methods could obtain completely correct classification.

參考資料

[1] 鄭華彬、袁久榮、袁洁，中成藥，2003，25(9)，附4。
[2] M. Twett, Proc. Warsaw Soc. Nat. Sci. Biol., 1903, 14, 6.
[3] A. J. P. Martin and R. L. M Synge, Biochem. J., 1941, 35, 1358.
[4] A. T. James and A. J. P. Martin, Analyst, 1952, 77, 915.
[5] E. Stahl, Chemiker-Ztg., 1958, 82, 323.
[6] J. F. K. Huber and J. A. R. J. Hulsman, Anal. Chim. Acta, 1967, 38, 305.
[7] 吳健三，高效液相層析儀實務篇，pp.30-33，建吾企業股份有限公司， 台北，1991。
[8] R. S. Houk, Anal. Chem. Res., 1994, 27, 333.
[9] R. S. Houk, V. A. Fassel, G. D. Flesch, H. J. Svec, A. L. Gray, C. E. Tayor, Anal. Chem., 1980, 52, 2283.
[10] R. S. Houk, Anal. Chem., 1986, 58, 97A.
[11] J. W. Olesik, Anal. Chem., 1991, 63, 12A.
[12] J. W. Olesik, Anal. Chem., 1996, 68, 469A
[13] R. Richard, M. J. Lazaro, H. Lachas and A. A. Herod, Rapid Commum. Mass Spectrom., 2000, 14, 317.
[14] H. Lachas, R. Richard, A. A. Herod, D. R. Dugwell, Rapid Commum. Mass Spectrom.,2000, 14, 335.
[15] 陳順宇，多變量分析，第一章，p.1，華太書局，台南，2004。
[16] 林清山，多變項統計分析法，東華書局，台北，1988。
[17] 陳順宇，多變量分析，第一章，p.5，華太書局，台南，2004。
[18] Tom A. B. Snijders, Roel J. Bosker, Multilevel analysis, pp.68-70, London,1993.
[19] 陳順宇，多變量分析，第一章，p.9，華太書局，台南，2004。
[20] 林清山，多變項統計分析法，p.40，東華書局，台北，1988。
[21] 林清山，多變項統計分析法，p.62，東華書局，台北，1988。
[22] 陳順宇，多變量分析，第三章，p.12，華太書局，台南，2004。
[23] 陳順宇，多變量分析，第三章，p.33，華太書局，台南，2004。
[24] 陳順宇，多變量分析，第四章，p.25，華太書局，台南，2004。
[25] 林清山，多變項統計分析法，p.77，東華書局，台北，1988。
[26] 林清山，多變項統計分析法，p.92，東華書局，台北，1988。
[27] 林清山，多變項統計分析法，p.142，東華書局，台北，1988。
[28] 林清山，多變項統計分析法，p.165，東華書局，台北，1988。