液相色谱氢化物发生原子荧光光谱法测定富硒酵母中硒的形态
肖志明 宋荣 贾铮 李阳 樊霞
摘要以不同厂家阿莫西林胶囊及其内容物近红外(Near infrared, NIR)光谱为例,寻找评价分段直接标准化算法(Piecewise direct standardization, PDS)进行光谱校正是否成功的量化指标。本研究共涉及76批阿莫西林胶囊样品,其中54批用于建立胶囊剂的定量模型。通过聚类分析,所有胶囊的NIR光谱分成5类,每类视为一个均质样本;分别计算每个均质样本的平均光谱,从该样本中选择10~15张光谱作为PDS校正的目标光谱,对76批阿莫西林胶囊内容物粉末光谱进行校正,利用阿莫西林胶囊定量模型对校正后的光谱进行含量预测;计算校正后的光谱与PDS校正中目标光谱所属均质样本的平均光谱的相似系数,分析其与预测误差的关系。结果表明,校正结果与所选择的目标光谱关系密切。PDS校正光谱与模型中不同均质样本平均光谱的相似系数(r)越大,通常校正效果越好;当r<99%时,一般可判断PDS校正失败(预测误差>5%)。因此, 可以用PDS校正后光谱与校正时使用的目标光谱所属的均质样本的平均光谱的相似系数作为判断PDS校正是否成功的标志。
关键词PDS算法; NIR定量模型; 预测结果; 误差分析
1引言
利用近红外(Near infrared, NIR)技术识别假劣药品和进行药品生产过程控制,已经成为药物分析的新热点\[1~4\]。NIR技术的应用与所采用的模型关系密切。NIR模型优劣不仅与建模所选择的谱段\[5,6\]、预处理方法\[7\]和算法\[6\]有关,更与建模训练集样本的代表性关系密切\[8,9\]。为表述NIR建模样本的代表性问题,本研究组提出了均质样本概念\[10,11\]。NIR定量模型的训练集中包含有若干个不同的均质样本;当模型遇到建模时未包括的新均质样本时,预测结果就可能出现较大偏差。这时可以通过加入新样本进行模型更新或利用化学计量学算法对新样本光谱进行校正,扩展原模型的适用范围\[10~13\]。为了解决NIR技术在企业生产过程控制应用之初代表性样品收集困难、建模繁琐问题,开展了对已建立的通用性模型经校正后作为生产过程控制初始模型的研究\[13\],已采用分段直接标准化算法(Piecewise direct standardization, PDS)和斜率/截距(Slope/Bias, S/B)算法,用头孢拉定胶囊定量模型直接预测生产过程的中间体胶囊内容物的含量,取得了良好的预测效果。但如何合理快速地判断这些校正方法的效果问题仍未解决。
本研究以阿莫西林胶囊定量模型为例,选择不同的目标光谱用于PDS校正,尝试将生产过程中间体阿莫西林胶囊内容物光谱经过PDS校正后通过阿莫西林胶囊定量模型预测含量。通过探讨PDS校正中,校正光谱和模型训练集中不同均质样本平均光谱的相似系数(简称校正光谱相似系数)与预测误差的关系,寻找用于判定PDS校正准确性的量化指标,为实际应用提供指导。
2方法原理
2.1光谱预处理方法
对光谱进行预处理可以提高定量分析中光谱数据与其对应含量值之间的相关关系。本实验主要采用的光谱预处理方法有:
2.1.1直线差减法直线差减法是一种针对倾斜光谱的经典基线校正方法。首先对校正的波段用最小二乘拟合一条直线,然后从光谱中减去该直线,达到基线校正的目的\[14\]。
2.1.2矢量归一化法在NIR固体样本的定量分析中,一般假设测量时近红外光在样品中的有效路程一致。但样品内部的粒度、晶型及测量重复性等因素都易引起测量光程的变化。
转换矩阵F就可以将转化光谱Xs转换成与目标光谱相匹配的光谱Xs,std。
利用PDS可以进行不同仪器光谱间系统误差的校正。本研究尝试将阿莫西林胶囊光谱(目标光谱)与其内容物粉末光谱(转化光谱)的差异看作是系统误差(主要为胶囊壳的差异),采用PDS法对胶囊内容物的光谱先进行校正,再利用阿莫西林胶囊含量预测模型预测胶囊内容物的含量。
2.3均质样本
均质样本是指主成分含量不同,辅料以及制剂工艺相同或相近的一组样品。该理念起源于通用性氧氟沙星注射液定量模型的研究\[11\],研究中发现模型训练集中,处方相同仅活性成分含量不同的样本在其主成分得分图中可明显集中于一组,该组样本被称为一个均质样本;辅料处方不同的样本属于不同的均质样本;均质样本的NIR光谱具有高度的相似性。后来该理念被延伸至药品固体制剂:认为NIR光谱具有高度相似性的一组样本为一个均质样本;均质样本可通过聚类分析的方法进行划分,用于建模时训练集样本的选择以及判断模型是否需要更新。均质样本光谱相似性的阈值可利用相关系数确定\[10\]。定量模型的训练集可以认为由一个或几个均质样本组成,建模时应从不同的均质样本中选择代表性样品组成训练集,当预测训练集未包含的均质样本样品时,预测误差变大,需要对模型更新或对该类样本进行校正。
3实验部分
3.1仪器与试剂
Bruker MatrixF傅立叶变换NIR光谱仪,配有光纤探头测样附件,铟镓砷(InGaAs)检测器,Bruker公司OPUS 5.5光谱分析软件。岛津20A高效液相色谱分析系统,配有自动进样器,二极管阵列检测器以及工作站。
76批阿莫西林胶囊(规格为0.25 g和0.50 g)为2010年全国评价性抽验样品,含量(mg/mg)范围为84.0%~67.5%;阿莫西林对照品(批号:130409201011),由中国食品药品检定研究院提供。
3.2含量参考值的测定
按中国药典2010版HPLC法测定阿莫西林含量\[18\]。色谱柱:Dikma Diamonsil C18 (250 mm×2.4 mm, 5 μm);流动相:0.05 mol/L KH2PO4溶液(用2 mol/L KOH调至pH 5.0)乙腈(97.5∶2.5, V/V);检测波长254 nm;柱温:30 ℃;流速:1.7 mL/min;进样量:20 μL。
3.3样品NIR光谱的采集
利用光纤探头分别采集阿莫西林胶囊和胶囊内容物光谱。光谱测量范围为4000~12000 cm
背景扫描次数为32次,样品扫描次数为32次,测定温度为室温(22±2)℃,湿度为20%~50%。
从每批样品中随机抽取6粒胶囊,将光纤抵在单层胶囊壳的一侧扫描,每粒扫描3张光谱,计算平均光谱。再将胶囊内容物分别倾倒至标准NIR测量瓶中,将光纤插入内容物中,扫描3张光谱,计算平均光谱。
3.4建立NIR模型
用Bruker OPUS软件中的Qunant 2模块,参照文献\[8,9\],采用PLS算法建立阿莫西林胶囊定量模型。按照参考文献\[7\]选择训练集样本:首先对所有的样品光谱经矢量归一化处理后在全谱范围内采用Wards算法进行聚类分析,从76批光谱中选择出54张光谱作为训练集,其它光谱作为验证集;建模谱段为5400~7100 cm
3.5PDS校正
利用OPUS 5.5软件中“Setup spectra transfer method”模块,采用PDS法对阿莫西林胶囊内容物的光谱进行校正,再利用所建立的阿莫西林胶囊定量模型预测胶囊内容物的含量。参照文献\[13\],设定PDS校正过程中使用的参数;目标光谱(阿莫西林胶囊光谱)数量一般选择10~15张,窗口大小选择7个波长点。
根据76批阿莫西林胶囊光谱的聚类分析结果,全部样本大致可分成5类,每一类被认为是一个均质样本。从5类光谱中分别选择PDS校正的目标光谱,分别称之为类Ⅰ、类Ⅱ、…、类Ⅴ光谱,计算同一胶囊内容物光谱经不同的目标光谱校正后得到的校正光谱与各均质样本平均光谱的相似系数
4结果与讨论
4.1阿莫西林胶囊定量模型
阿莫西林胶囊、胶囊内容物和胶囊壳的NIR光谱呈明显差异
4.2PDS校正
由于胶囊内容物光谱与阿莫西林胶囊光谱具有较大的差异,直接利用阿莫西林胶囊模型预测胶囊内容物的含量,误差均大于5%;经PDS校正后可以改善预测的准确性,但部分样本的预测误差较大(>5%)(表2)。
PDS校正系通过对胶囊光谱和对应的胶
囊内容物光谱进行关联,将内容物光谱校正成胶囊光谱进行预测。分别用类Ⅰ、类Ⅱ、…、类Ⅴ光谱对每一个内容物光谱进行PDS校正(得到的校正光谱分别称类Ⅰ、类Ⅱ、…、类Ⅴ校正光谱),预测含量,计算预测误差;再计算诸校正光谱与阿莫西林胶囊各均质样本平均光谱的相似系数,简称校正光谱相似系数(r);将每一个校正光谱的预测误差与对应的r值作图,以预测误差5%为分界,分析预测误差与r的关系。
分析类Ⅰ校正光谱的预测误差和与之对应的r之间的关系,发现预测误差随着r的增大而减小;误差大于5%的样本共有14个,其中r最大为98.78%,最小为93.92%;误差小于5%的样本共有62个,其中r最大值为99.87%,最小值为98.84%。由主成分得分图可见(图2),预测误差大于5%的光谱均分布在训练集样本范围之外,说明其与训练集光谱存在较大的差异;预测误差小于2%的样本,则基本都分布在训练集范围之内,说明其与训练集光谱的相似性较高;证明了校正光谱的相似性直接影响预测结果。
同法分别分析类Ⅱ、…、类Ⅴ校正光谱的预测误差和其在模型训练集光谱主成分得分图中的位置(图3),结果均与类Ⅰ校正光谱相似,预测误差随着r的增大而减小。预测误差和r汇总于表3。结果表明,当PDS校正光谱与建模的均质样本光谱均存在较大差异时,模型将不能对其准确预测。
绘制r的正态分布曲线(图4),单侧检验,计算出其95%的置信区间为0.9863~0.9994。由于此正态分布中的r所对应的校正光谱的预测误差均小于5%,故可以认为当r>98.63%时,模型对校正光谱的预测误差小于5%。即r=99%可作为阈值, 用于判断PDS校准成功与否。5结论
PDS校正可以扩展NIR模型的适用范围,但校正的成功与否与所选择的目标光谱关系密切。利用均质样本概念,计算PDS校正光谱与模型中诸均质样本光谱的相似系数(r)。通常r越大,校正效果越好;当r<99%时,一般可判断PDS校正失败(预测误差>5%),据此可以选择适宜的目标光谱进行PDS校正,也可以判断PDS校正的成功与否。
References
1Moffat A C, Assi S, Watt R A. J. Near Infrared Spec., 2010, 18(1): 1-15
2FENG YanChun, HU ChangQin. Spectrosc. Spect. Anal., 2014, 34(5): 1222-1228
冯艳春, 胡昌勤. 光谱学与光谱分析, 2014, 34(5): 1222-1228
3Reich G. Adv. Drug Deliver. Rev., 2005, 57: 1109-1143
4Boyer C, Gaudin K, Kauss T, Gaubert A, Boudis A, Verschelden J, Franca M, Roussille J, Boucher J, Olliaro P, White N J, Millet P, Dubost J. J. Pharm. Biomed. Anal., 2012, 6768: 10-15
5Li H D, Liang Y Z, Xu Q S, Cao D S. Anal. Chim. Acta, 2009, 648: 77-84
6Liu Y Y, Hu C Q. Vib. Spectrosc., 2011, 55: 241-249
7Ni Z, Feng Y C, Hu C Q. J. Anal. Bioanal. Techniques, 2010, 1(3): 1-7
8Jia Y H, Liu X P, Feng Y C, Hu C Q. AAPS PharmSciTech., 2011, 12(2): 738-745
9Zou W B, Feng Y C, Song D Q, Hu C Q. J. Chin. Pharm. Sci., 2012, 21(5): 459-467
10Zou W B, Feng Y C, Dong J X, Song D Q, Hu C Q. Sci. China. Chem., 2013, 56(4): 533-540
11Hou S R, Feng Y C, Zhang X B, Hu C Q. J. Chin. Pharm. Sci., 2012, 21(1): 62-69
12Zhang X B, Feng Y C, Hu C Q. Anal. Chim. Acta, 2008, 630: 131-140
13LEI DeQing, HU ChangQin, FENG YanChun, FENG Fang. Acta Pharm. Sin., 2010, 45 (11): 1421-1426
雷德卿, 胡昌勤, 冯艳春, 冯 芳. 药学学报, 2010, 45 (11): 1421-1426
14KE BoKe. Research on Application of Near Infrared Spectroscopy in Silicagel Column Chromatography Process. Hangzhou: Zhejiang University, 2007: 8
柯博克. 近红外光谱在硅胶柱层析过程分析中的应用研究. 杭州: 浙江大学, 2007: 8
15NI Zhen,HU ChangQin, FENG Fang. Chin. J. Pharm. Anal., 2008, 28(5): 824-829
尼 珍, 胡昌勤, 冯 芳. 药物分析杂志, 2008, 28(5): 824-829
16Candolfi A, Maesschalck De R, JouanRimbaud D, Hailey P A, Massart D L. J. Pharm. Biomed. Anal., 1999, 21: 115-132
17ZHANG XueBo, FENG YanChun, HU ChangQin. Chin. J. Pharm. Anal., 2009, 29(8): 1390-1399
张学博, 冯艳春, 胡昌勤. 药物分析杂志, 2009, 29(8): 1390-1399
18Chinese Pharmacopoeia Commission. Chinese Pharmacopoeia. Beijing: Chinese Medical Science and Technology Press, 2010, 401-402
国家药典委员会. 中国药典. 北京: 中国医药科技出版社, 2010: 401-402
AbstractThe near infrared (NIR) spectra of 76 batches of the amoxicillin capsules from different manufacturers and their corresponding content powder without capsules cell were used to find some quantitative indicators to evaluate whether the piecewise direct standardization (PDS) algorithm succeeded in NIR quantitative model updating. 54 batches were used to construct the NIR quantitative model for capsule preparation. All the NIR spectra of amoxicillin capsules were divided into five classes by cluster analysis, and each class can be regarded as a homology sample set. The average spectrum for each homology sample set was calculated. Ten to Fifteen spectra were selected from each homology sample set as the corresponding master spectra of the PDS algorithm to correct all the NIR spectra of the amoxicillin content powder respectively. Then the corrected spectra were predicted by the constructed NIR quantitative model for amoxicillin capsules. The prediction error for each corrected powder spectrum, and the correlation coefficient between each corrected powder spectrum and the average spectrum of the corresponding homolog sample set which the PDS master spectra came from, were calculated. Finally, the relationship between the prediction error and its corresponding correlation coefficient were studied. It was found that the correction results correlated closely with the selected master spectra set in PDS algorithm. The bigger the correlation coefficient (r), the better the correction results. In general, when r is less than 99%, it can be judged that the PDS correction is failed. At this condition, the prediction error is often more than 5%. Therefore, the correlation coefficient between the corrected spectrum and its corresponding average spectrum of the homology sample set can be used as an indicator to evaluate the efficiency of the PDS correction.
KeywordsPiecewise direct standardization algorithm; Near infrared quantitative model; Prediction results; Error analysis
7Ni Z, Feng Y C, Hu C Q. J. Anal. Bioanal. Techniques, 2010, 1(3): 1-7
8Jia Y H, Liu X P, Feng Y C, Hu C Q. AAPS PharmSciTech., 2011, 12(2): 738-745
9Zou W B, Feng Y C, Song D Q, Hu C Q. J. Chin. Pharm. Sci., 2012, 21(5): 459-467
10Zou W B, Feng Y C, Dong J X, Song D Q, Hu C Q. Sci. China. Chem., 2013, 56(4): 533-540
11Hou S R, Feng Y C, Zhang X B, Hu C Q. J. Chin. Pharm. Sci., 2012, 21(1): 62-69
12Zhang X B, Feng Y C, Hu C Q. Anal. Chim. Acta, 2008, 630: 131-140
13LEI DeQing, HU ChangQin, FENG YanChun, FENG Fang. Acta Pharm. Sin., 2010, 45 (11): 1421-1426
雷德卿, 胡昌勤, 冯艳春, 冯 芳. 药学学报, 2010, 45 (11): 1421-1426
14KE BoKe. Research on Application of Near Infrared Spectroscopy in Silicagel Column Chromatography Process. Hangzhou: Zhejiang University, 2007: 8
柯博克. 近红外光谱在硅胶柱层析过程分析中的应用研究. 杭州: 浙江大学, 2007: 8
15NI Zhen,HU ChangQin, FENG Fang. Chin. J. Pharm. Anal., 2008, 28(5): 824-829
尼 珍, 胡昌勤, 冯 芳. 药物分析杂志, 2008, 28(5): 824-829
16Candolfi A, Maesschalck De R, JouanRimbaud D, Hailey P A, Massart D L. J. Pharm. Biomed. Anal., 1999, 21: 115-132
17ZHANG XueBo, FENG YanChun, HU ChangQin. Chin. J. Pharm. Anal., 2009, 29(8): 1390-1399
张学博, 冯艳春, 胡昌勤. 药物分析杂志, 2009, 29(8): 1390-1399
18Chinese Pharmacopoeia Commission. Chinese Pharmacopoeia. Beijing: Chinese Medical Science and Technology Press, 2010, 401-402
国家药典委员会. 中国药典. 北京: 中国医药科技出版社, 2010: 401-402
AbstractThe near infrared (NIR) spectra of 76 batches of the amoxicillin capsules from different manufacturers and their corresponding content powder without capsules cell were used to find some quantitative indicators to evaluate whether the piecewise direct standardization (PDS) algorithm succeeded in NIR quantitative model updating. 54 batches were used to construct the NIR quantitative model for capsule preparation. All the NIR spectra of amoxicillin capsules were divided into five classes by cluster analysis, and each class can be regarded as a homology sample set. The average spectrum for each homology sample set was calculated. Ten to Fifteen spectra were selected from each homology sample set as the corresponding master spectra of the PDS algorithm to correct all the NIR spectra of the amoxicillin content powder respectively. Then the corrected spectra were predicted by the constructed NIR quantitative model for amoxicillin capsules. The prediction error for each corrected powder spectrum, and the correlation coefficient between each corrected powder spectrum and the average spectrum of the corresponding homolog sample set which the PDS master spectra came from, were calculated. Finally, the relationship between the prediction error and its corresponding correlation coefficient were studied. It was found that the correction results correlated closely with the selected master spectra set in PDS algorithm. The bigger the correlation coefficient (r), the better the correction results. In general, when r is less than 99%, it can be judged that the PDS correction is failed. At this condition, the prediction error is often more than 5%. Therefore, the correlation coefficient between the corrected spectrum and its corresponding average spectrum of the homology sample set can be used as an indicator to evaluate the efficiency of the PDS correction.
KeywordsPiecewise direct standardization algorithm; Near infrared quantitative model; Prediction results; Error analysis
7Ni Z, Feng Y C, Hu C Q. J. Anal. Bioanal. Techniques, 2010, 1(3): 1-7
8Jia Y H, Liu X P, Feng Y C, Hu C Q. AAPS PharmSciTech., 2011, 12(2): 738-745
9Zou W B, Feng Y C, Song D Q, Hu C Q. J. Chin. Pharm. Sci., 2012, 21(5): 459-467
10Zou W B, Feng Y C, Dong J X, Song D Q, Hu C Q. Sci. China. Chem., 2013, 56(4): 533-540
11Hou S R, Feng Y C, Zhang X B, Hu C Q. J. Chin. Pharm. Sci., 2012, 21(1): 62-69
12Zhang X B, Feng Y C, Hu C Q. Anal. Chim. Acta, 2008, 630: 131-140
13LEI DeQing, HU ChangQin, FENG YanChun, FENG Fang. Acta Pharm. Sin., 2010, 45 (11): 1421-1426
雷德卿, 胡昌勤, 冯艳春, 冯 芳. 药学学报, 2010, 45 (11): 1421-1426
14KE BoKe. Research on Application of Near Infrared Spectroscopy in Silicagel Column Chromatography Process. Hangzhou: Zhejiang University, 2007: 8
柯博克. 近红外光谱在硅胶柱层析过程分析中的应用研究. 杭州: 浙江大学, 2007: 8
15NI Zhen,HU ChangQin, FENG Fang. Chin. J. Pharm. Anal., 2008, 28(5): 824-829
尼 珍, 胡昌勤, 冯 芳. 药物分析杂志, 2008, 28(5): 824-829
16Candolfi A, Maesschalck De R, JouanRimbaud D, Hailey P A, Massart D L. J. Pharm. Biomed. Anal., 1999, 21: 115-132
17ZHANG XueBo, FENG YanChun, HU ChangQin. Chin. J. Pharm. Anal., 2009, 29(8): 1390-1399
张学博, 冯艳春, 胡昌勤. 药物分析杂志, 2009, 29(8): 1390-1399
18Chinese Pharmacopoeia Commission. Chinese Pharmacopoeia. Beijing: Chinese Medical Science and Technology Press, 2010, 401-402
国家药典委员会. 中国药典. 北京: 中国医药科技出版社, 2010: 401-402
AbstractThe near infrared (NIR) spectra of 76 batches of the amoxicillin capsules from different manufacturers and their corresponding content powder without capsules cell were used to find some quantitative indicators to evaluate whether the piecewise direct standardization (PDS) algorithm succeeded in NIR quantitative model updating. 54 batches were used to construct the NIR quantitative model for capsule preparation. All the NIR spectra of amoxicillin capsules were divided into five classes by cluster analysis, and each class can be regarded as a homology sample set. The average spectrum for each homology sample set was calculated. Ten to Fifteen spectra were selected from each homology sample set as the corresponding master spectra of the PDS algorithm to correct all the NIR spectra of the amoxicillin content powder respectively. Then the corrected spectra were predicted by the constructed NIR quantitative model for amoxicillin capsules. The prediction error for each corrected powder spectrum, and the correlation coefficient between each corrected powder spectrum and the average spectrum of the corresponding homolog sample set which the PDS master spectra came from, were calculated. Finally, the relationship between the prediction error and its corresponding correlation coefficient were studied. It was found that the correction results correlated closely with the selected master spectra set in PDS algorithm. The bigger the correlation coefficient (r), the better the correction results. In general, when r is less than 99%, it can be judged that the PDS correction is failed. At this condition, the prediction error is often more than 5%. Therefore, the correlation coefficient between the corrected spectrum and its corresponding average spectrum of the homology sample set can be used as an indicator to evaluate the efficiency of the PDS correction.
KeywordsPiecewise direct standardization algorithm; Near infrared quantitative model; Prediction results; Error analysis
摘要以不同厂家阿莫西林胶囊及其内容物近红外(Near infrared, NIR)光谱为例,寻找评价分段直接标准化算法(Piecewise direct standardization, PDS)进行光谱校正是否成功的量化指标。本研究共涉及76批阿莫西林胶囊样品,其中54批用于建立胶囊剂的定量模型。通过聚类分析,所有胶囊的NIR光谱分成5类,每类视为一个均质样本;分别计算每个均质样本的平均光谱,从该样本中选择10~15张光谱作为PDS校正的目标光谱,对76批阿莫西林胶囊内容物粉末光谱进行校正,利用阿莫西林胶囊定量模型对校正后的光谱进行含量预测;计算校正后的光谱与PDS校正中目标光谱所属均质样本的平均光谱的相似系数,分析其与预测误差的关系。结果表明,校正结果与所选择的目标光谱关系密切。PDS校正光谱与模型中不同均质样本平均光谱的相似系数(r)越大,通常校正效果越好;当r<99%时,一般可判断PDS校正失败(预测误差>5%)。因此, 可以用PDS校正后光谱与校正时使用的目标光谱所属的均质样本的平均光谱的相似系数作为判断PDS校正是否成功的标志。
关键词PDS算法; NIR定量模型; 预测结果; 误差分析
1引言
利用近红外(Near infrared, NIR)技术识别假劣药品和进行药品生产过程控制,已经成为药物分析的新热点\[1~4\]。NIR技术的应用与所采用的模型关系密切。NIR模型优劣不仅与建模所选择的谱段\[5,6\]、预处理方法\[7\]和算法\[6\]有关,更与建模训练集样本的代表性关系密切\[8,9\]。为表述NIR建模样本的代表性问题,本研究组提出了均质样本概念\[10,11\]。NIR定量模型的训练集中包含有若干个不同的均质样本;当模型遇到建模时未包括的新均质样本时,预测结果就可能出现较大偏差。这时可以通过加入新样本进行模型更新或利用化学计量学算法对新样本光谱进行校正,扩展原模型的适用范围\[10~13\]。为了解决NIR技术在企业生产过程控制应用之初代表性样品收集困难、建模繁琐问题,开展了对已建立的通用性模型经校正后作为生产过程控制初始模型的研究\[13\],已采用分段直接标准化算法(Piecewise direct standardization, PDS)和斜率/截距(Slope/Bias, S/B)算法,用头孢拉定胶囊定量模型直接预测生产过程的中间体胶囊内容物的含量,取得了良好的预测效果。但如何合理快速地判断这些校正方法的效果问题仍未解决。
本研究以阿莫西林胶囊定量模型为例,选择不同的目标光谱用于PDS校正,尝试将生产过程中间体阿莫西林胶囊内容物光谱经过PDS校正后通过阿莫西林胶囊定量模型预测含量。通过探讨PDS校正中,校正光谱和模型训练集中不同均质样本平均光谱的相似系数(简称校正光谱相似系数)与预测误差的关系,寻找用于判定PDS校正准确性的量化指标,为实际应用提供指导。
2方法原理
2.1光谱预处理方法
对光谱进行预处理可以提高定量分析中光谱数据与其对应含量值之间的相关关系。本实验主要采用的光谱预处理方法有:
2.1.1直线差减法直线差减法是一种针对倾斜光谱的经典基线校正方法。首先对校正的波段用最小二乘拟合一条直线,然后从光谱中减去该直线,达到基线校正的目的\[14\]。
2.1.2矢量归一化法在NIR固体样本的定量分析中,一般假设测量时近红外光在样品中的有效路程一致。但样品内部的粒度、晶型及测量重复性等因素都易引起测量光程的变化。
转换矩阵F就可以将转化光谱Xs转换成与目标光谱相匹配的光谱Xs,std。
利用PDS可以进行不同仪器光谱间系统误差的校正。本研究尝试将阿莫西林胶囊光谱(目标光谱)与其内容物粉末光谱(转化光谱)的差异看作是系统误差(主要为胶囊壳的差异),采用PDS法对胶囊内容物的光谱先进行校正,再利用阿莫西林胶囊含量预测模型预测胶囊内容物的含量。
2.3均质样本
均质样本是指主成分含量不同,辅料以及制剂工艺相同或相近的一组样品。该理念起源于通用性氧氟沙星注射液定量模型的研究\[11\],研究中发现模型训练集中,处方相同仅活性成分含量不同的样本在其主成分得分图中可明显集中于一组,该组样本被称为一个均质样本;辅料处方不同的样本属于不同的均质样本;均质样本的NIR光谱具有高度的相似性。后来该理念被延伸至药品固体制剂:认为NIR光谱具有高度相似性的一组样本为一个均质样本;均质样本可通过聚类分析的方法进行划分,用于建模时训练集样本的选择以及判断模型是否需要更新。均质样本光谱相似性的阈值可利用相关系数确定\[10\]。定量模型的训练集可以认为由一个或几个均质样本组成,建模时应从不同的均质样本中选择代表性样品组成训练集,当预测训练集未包含的均质样本样品时,预测误差变大,需要对模型更新或对该类样本进行校正。
3实验部分
3.1仪器与试剂
Bruker MatrixF傅立叶变换NIR光谱仪,配有光纤探头测样附件,铟镓砷(InGaAs)检测器,Bruker公司OPUS 5.5光谱分析软件。岛津20A高效液相色谱分析系统,配有自动进样器,二极管阵列检测器以及工作站。
76批阿莫西林胶囊(规格为0.25 g和0.50 g)为2010年全国评价性抽验样品,含量(mg/mg)范围为84.0%~67.5%;阿莫西林对照品(批号:130409201011),由中国食品药品检定研究院提供。
3.2含量参考值的测定
按中国药典2010版HPLC法测定阿莫西林含量\[18\]。色谱柱:Dikma Diamonsil C18 (250 mm×2.4 mm, 5 μm);流动相:0.05 mol/L KH2PO4溶液(用2 mol/L KOH调至pH 5.0)乙腈(97.5∶2.5, V/V);检测波长254 nm;柱温:30 ℃;流速:1.7 mL/min;进样量:20 μL。
3.3样品NIR光谱的采集
利用光纤探头分别采集阿莫西林胶囊和胶囊内容物光谱。光谱测量范围为4000~12000 cm
背景扫描次数为32次,样品扫描次数为32次,测定温度为室温(22±2)℃,湿度为20%~50%。
从每批样品中随机抽取6粒胶囊,将光纤抵在单层胶囊壳的一侧扫描,每粒扫描3张光谱,计算平均光谱。再将胶囊内容物分别倾倒至标准NIR测量瓶中,将光纤插入内容物中,扫描3张光谱,计算平均光谱。
3.4建立NIR模型
用Bruker OPUS软件中的Qunant 2模块,参照文献\[8,9\],采用PLS算法建立阿莫西林胶囊定量模型。按照参考文献\[7\]选择训练集样本:首先对所有的样品光谱经矢量归一化处理后在全谱范围内采用Wards算法进行聚类分析,从76批光谱中选择出54张光谱作为训练集,其它光谱作为验证集;建模谱段为5400~7100 cm
3.5PDS校正
利用OPUS 5.5软件中“Setup spectra transfer method”模块,采用PDS法对阿莫西林胶囊内容物的光谱进行校正,再利用所建立的阿莫西林胶囊定量模型预测胶囊内容物的含量。参照文献\[13\],设定PDS校正过程中使用的参数;目标光谱(阿莫西林胶囊光谱)数量一般选择10~15张,窗口大小选择7个波长点。
根据76批阿莫西林胶囊光谱的聚类分析结果,全部样本大致可分成5类,每一类被认为是一个均质样本。从5类光谱中分别选择PDS校正的目标光谱,分别称之为类Ⅰ、类Ⅱ、…、类Ⅴ光谱,计算同一胶囊内容物光谱经不同的目标光谱校正后得到的校正光谱与各均质样本平均光谱的相似系数
4结果与讨论
4.1阿莫西林胶囊定量模型
阿莫西林胶囊、胶囊内容物和胶囊壳的NIR光谱呈明显差异
4.2PDS校正
由于胶囊内容物光谱与阿莫西林胶囊光谱具有较大的差异,直接利用阿莫西林胶囊模型预测胶囊内容物的含量,误差均大于5%;经PDS校正后可以改善预测的准确性,但部分样本的预测误差较大(>5%)(表2)。
PDS校正系通过对胶囊光谱和对应的胶
囊内容物光谱进行关联,将内容物光谱校正成胶囊光谱进行预测。分别用类Ⅰ、类Ⅱ、…、类Ⅴ光谱对每一个内容物光谱进行PDS校正(得到的校正光谱分别称类Ⅰ、类Ⅱ、…、类Ⅴ校正光谱),预测含量,计算预测误差;再计算诸校正光谱与阿莫西林胶囊各均质样本平均光谱的相似系数,简称校正光谱相似系数(r);将每一个校正光谱的预测误差与对应的r值作图,以预测误差5%为分界,分析预测误差与r的关系。
分析类Ⅰ校正光谱的预测误差和与之对应的r之间的关系,发现预测误差随着r的增大而减小;误差大于5%的样本共有14个,其中r最大为98.78%,最小为93.92%;误差小于5%的样本共有62个,其中r最大值为99.87%,最小值为98.84%。由主成分得分图可见(图2),预测误差大于5%的光谱均分布在训练集样本范围之外,说明其与训练集光谱存在较大的差异;预测误差小于2%的样本,则基本都分布在训练集范围之内,说明其与训练集光谱的相似性较高;证明了校正光谱的相似性直接影响预测结果。
同法分别分析类Ⅱ、…、类Ⅴ校正光谱的预测误差和其在模型训练集光谱主成分得分图中的位置(图3),结果均与类Ⅰ校正光谱相似,预测误差随着r的增大而减小。预测误差和r汇总于表3。结果表明,当PDS校正光谱与建模的均质样本光谱均存在较大差异时,模型将不能对其准确预测。
绘制r的正态分布曲线(图4),单侧检验,计算出其95%的置信区间为0.9863~0.9994。由于此正态分布中的r所对应的校正光谱的预测误差均小于5%,故可以认为当r>98.63%时,模型对校正光谱的预测误差小于5%。即r=99%可作为阈值, 用于判断PDS校准成功与否。5结论
PDS校正可以扩展NIR模型的适用范围,但校正的成功与否与所选择的目标光谱关系密切。利用均质样本概念,计算PDS校正光谱与模型中诸均质样本光谱的相似系数(r)。通常r越大,校正效果越好;当r<99%时,一般可判断PDS校正失败(预测误差>5%),据此可以选择适宜的目标光谱进行PDS校正,也可以判断PDS校正的成功与否。
References
1Moffat A C, Assi S, Watt R A. J. Near Infrared Spec., 2010, 18(1): 1-15
2FENG YanChun, HU ChangQin. Spectrosc. Spect. Anal., 2014, 34(5): 1222-1228
冯艳春, 胡昌勤. 光谱学与光谱分析, 2014, 34(5): 1222-1228
3Reich G. Adv. Drug Deliver. Rev., 2005, 57: 1109-1143
4Boyer C, Gaudin K, Kauss T, Gaubert A, Boudis A, Verschelden J, Franca M, Roussille J, Boucher J, Olliaro P, White N J, Millet P, Dubost J. J. Pharm. Biomed. Anal., 2012, 6768: 10-15
5Li H D, Liang Y Z, Xu Q S, Cao D S. Anal. Chim. Acta, 2009, 648: 77-84
6Liu Y Y, Hu C Q. Vib. Spectrosc., 2011, 55: 241-249
7Ni Z, Feng Y C, Hu C Q. J. Anal. Bioanal. Techniques, 2010, 1(3): 1-7
8Jia Y H, Liu X P, Feng Y C, Hu C Q. AAPS PharmSciTech., 2011, 12(2): 738-745
9Zou W B, Feng Y C, Song D Q, Hu C Q. J. Chin. Pharm. Sci., 2012, 21(5): 459-467
10Zou W B, Feng Y C, Dong J X, Song D Q, Hu C Q. Sci. China. Chem., 2013, 56(4): 533-540
11Hou S R, Feng Y C, Zhang X B, Hu C Q. J. Chin. Pharm. Sci., 2012, 21(1): 62-69
12Zhang X B, Feng Y C, Hu C Q. Anal. Chim. Acta, 2008, 630: 131-140
13LEI DeQing, HU ChangQin, FENG YanChun, FENG Fang. Acta Pharm. Sin., 2010, 45 (11): 1421-1426
雷德卿, 胡昌勤, 冯艳春, 冯 芳. 药学学报, 2010, 45 (11): 1421-1426
14KE BoKe. Research on Application of Near Infrared Spectroscopy in Silicagel Column Chromatography Process. Hangzhou: Zhejiang University, 2007: 8
柯博克. 近红外光谱在硅胶柱层析过程分析中的应用研究. 杭州: 浙江大学, 2007: 8
15NI Zhen,HU ChangQin, FENG Fang. Chin. J. Pharm. Anal., 2008, 28(5): 824-829
尼 珍, 胡昌勤, 冯 芳. 药物分析杂志, 2008, 28(5): 824-829
16Candolfi A, Maesschalck De R, JouanRimbaud D, Hailey P A, Massart D L. J. Pharm. Biomed. Anal., 1999, 21: 115-132
17ZHANG XueBo, FENG YanChun, HU ChangQin. Chin. J. Pharm. Anal., 2009, 29(8): 1390-1399
张学博, 冯艳春, 胡昌勤. 药物分析杂志, 2009, 29(8): 1390-1399
18Chinese Pharmacopoeia Commission. Chinese Pharmacopoeia. Beijing: Chinese Medical Science and Technology Press, 2010, 401-402
国家药典委员会. 中国药典. 北京: 中国医药科技出版社, 2010: 401-402
AbstractThe near infrared (NIR) spectra of 76 batches of the amoxicillin capsules from different manufacturers and their corresponding content powder without capsules cell were used to find some quantitative indicators to evaluate whether the piecewise direct standardization (PDS) algorithm succeeded in NIR quantitative model updating. 54 batches were used to construct the NIR quantitative model for capsule preparation. All the NIR spectra of amoxicillin capsules were divided into five classes by cluster analysis, and each class can be regarded as a homology sample set. The average spectrum for each homology sample set was calculated. Ten to Fifteen spectra were selected from each homology sample set as the corresponding master spectra of the PDS algorithm to correct all the NIR spectra of the amoxicillin content powder respectively. Then the corrected spectra were predicted by the constructed NIR quantitative model for amoxicillin capsules. The prediction error for each corrected powder spectrum, and the correlation coefficient between each corrected powder spectrum and the average spectrum of the corresponding homolog sample set which the PDS master spectra came from, were calculated. Finally, the relationship between the prediction error and its corresponding correlation coefficient were studied. It was found that the correction results correlated closely with the selected master spectra set in PDS algorithm. The bigger the correlation coefficient (r), the better the correction results. In general, when r is less than 99%, it can be judged that the PDS correction is failed. At this condition, the prediction error is often more than 5%. Therefore, the correlation coefficient between the corrected spectrum and its corresponding average spectrum of the homology sample set can be used as an indicator to evaluate the efficiency of the PDS correction.
KeywordsPiecewise direct standardization algorithm; Near infrared quantitative model; Prediction results; Error analysis
7Ni Z, Feng Y C, Hu C Q. J. Anal. Bioanal. Techniques, 2010, 1(3): 1-7
8Jia Y H, Liu X P, Feng Y C, Hu C Q. AAPS PharmSciTech., 2011, 12(2): 738-745
9Zou W B, Feng Y C, Song D Q, Hu C Q. J. Chin. Pharm. Sci., 2012, 21(5): 459-467
10Zou W B, Feng Y C, Dong J X, Song D Q, Hu C Q. Sci. China. Chem., 2013, 56(4): 533-540
11Hou S R, Feng Y C, Zhang X B, Hu C Q. J. Chin. Pharm. Sci., 2012, 21(1): 62-69
12Zhang X B, Feng Y C, Hu C Q. Anal. Chim. Acta, 2008, 630: 131-140
13LEI DeQing, HU ChangQin, FENG YanChun, FENG Fang. Acta Pharm. Sin., 2010, 45 (11): 1421-1426
雷德卿, 胡昌勤, 冯艳春, 冯 芳. 药学学报, 2010, 45 (11): 1421-1426
14KE BoKe. Research on Application of Near Infrared Spectroscopy in Silicagel Column Chromatography Process. Hangzhou: Zhejiang University, 2007: 8
柯博克. 近红外光谱在硅胶柱层析过程分析中的应用研究. 杭州: 浙江大学, 2007: 8
15NI Zhen,HU ChangQin, FENG Fang. Chin. J. Pharm. Anal., 2008, 28(5): 824-829
尼 珍, 胡昌勤, 冯 芳. 药物分析杂志, 2008, 28(5): 824-829
16Candolfi A, Maesschalck De R, JouanRimbaud D, Hailey P A, Massart D L. J. Pharm. Biomed. Anal., 1999, 21: 115-132
17ZHANG XueBo, FENG YanChun, HU ChangQin. Chin. J. Pharm. Anal., 2009, 29(8): 1390-1399
张学博, 冯艳春, 胡昌勤. 药物分析杂志, 2009, 29(8): 1390-1399
18Chinese Pharmacopoeia Commission. Chinese Pharmacopoeia. Beijing: Chinese Medical Science and Technology Press, 2010, 401-402
国家药典委员会. 中国药典. 北京: 中国医药科技出版社, 2010: 401-402
AbstractThe near infrared (NIR) spectra of 76 batches of the amoxicillin capsules from different manufacturers and their corresponding content powder without capsules cell were used to find some quantitative indicators to evaluate whether the piecewise direct standardization (PDS) algorithm succeeded in NIR quantitative model updating. 54 batches were used to construct the NIR quantitative model for capsule preparation. All the NIR spectra of amoxicillin capsules were divided into five classes by cluster analysis, and each class can be regarded as a homology sample set. The average spectrum for each homology sample set was calculated. Ten to Fifteen spectra were selected from each homology sample set as the corresponding master spectra of the PDS algorithm to correct all the NIR spectra of the amoxicillin content powder respectively. Then the corrected spectra were predicted by the constructed NIR quantitative model for amoxicillin capsules. The prediction error for each corrected powder spectrum, and the correlation coefficient between each corrected powder spectrum and the average spectrum of the corresponding homolog sample set which the PDS master spectra came from, were calculated. Finally, the relationship between the prediction error and its corresponding correlation coefficient were studied. It was found that the correction results correlated closely with the selected master spectra set in PDS algorithm. The bigger the correlation coefficient (r), the better the correction results. In general, when r is less than 99%, it can be judged that the PDS correction is failed. At this condition, the prediction error is often more than 5%. Therefore, the correlation coefficient between the corrected spectrum and its corresponding average spectrum of the homology sample set can be used as an indicator to evaluate the efficiency of the PDS correction.
KeywordsPiecewise direct standardization algorithm; Near infrared quantitative model; Prediction results; Error analysis
7Ni Z, Feng Y C, Hu C Q. J. Anal. Bioanal. Techniques, 2010, 1(3): 1-7
8Jia Y H, Liu X P, Feng Y C, Hu C Q. AAPS PharmSciTech., 2011, 12(2): 738-745
9Zou W B, Feng Y C, Song D Q, Hu C Q. J. Chin. Pharm. Sci., 2012, 21(5): 459-467
10Zou W B, Feng Y C, Dong J X, Song D Q, Hu C Q. Sci. China. Chem., 2013, 56(4): 533-540
11Hou S R, Feng Y C, Zhang X B, Hu C Q. J. Chin. Pharm. Sci., 2012, 21(1): 62-69
12Zhang X B, Feng Y C, Hu C Q. Anal. Chim. Acta, 2008, 630: 131-140
13LEI DeQing, HU ChangQin, FENG YanChun, FENG Fang. Acta Pharm. Sin., 2010, 45 (11): 1421-1426
雷德卿, 胡昌勤, 冯艳春, 冯 芳. 药学学报, 2010, 45 (11): 1421-1426
14KE BoKe. Research on Application of Near Infrared Spectroscopy in Silicagel Column Chromatography Process. Hangzhou: Zhejiang University, 2007: 8
柯博克. 近红外光谱在硅胶柱层析过程分析中的应用研究. 杭州: 浙江大学, 2007: 8
15NI Zhen,HU ChangQin, FENG Fang. Chin. J. Pharm. Anal., 2008, 28(5): 824-829
尼 珍, 胡昌勤, 冯 芳. 药物分析杂志, 2008, 28(5): 824-829
16Candolfi A, Maesschalck De R, JouanRimbaud D, Hailey P A, Massart D L. J. Pharm. Biomed. Anal., 1999, 21: 115-132
17ZHANG XueBo, FENG YanChun, HU ChangQin. Chin. J. Pharm. Anal., 2009, 29(8): 1390-1399
张学博, 冯艳春, 胡昌勤. 药物分析杂志, 2009, 29(8): 1390-1399
18Chinese Pharmacopoeia Commission. Chinese Pharmacopoeia. Beijing: Chinese Medical Science and Technology Press, 2010, 401-402
国家药典委员会. 中国药典. 北京: 中国医药科技出版社, 2010: 401-402
AbstractThe near infrared (NIR) spectra of 76 batches of the amoxicillin capsules from different manufacturers and their corresponding content powder without capsules cell were used to find some quantitative indicators to evaluate whether the piecewise direct standardization (PDS) algorithm succeeded in NIR quantitative model updating. 54 batches were used to construct the NIR quantitative model for capsule preparation. All the NIR spectra of amoxicillin capsules were divided into five classes by cluster analysis, and each class can be regarded as a homology sample set. The average spectrum for each homology sample set was calculated. Ten to Fifteen spectra were selected from each homology sample set as the corresponding master spectra of the PDS algorithm to correct all the NIR spectra of the amoxicillin content powder respectively. Then the corrected spectra were predicted by the constructed NIR quantitative model for amoxicillin capsules. The prediction error for each corrected powder spectrum, and the correlation coefficient between each corrected powder spectrum and the average spectrum of the corresponding homolog sample set which the PDS master spectra came from, were calculated. Finally, the relationship between the prediction error and its corresponding correlation coefficient were studied. It was found that the correction results correlated closely with the selected master spectra set in PDS algorithm. The bigger the correlation coefficient (r), the better the correction results. In general, when r is less than 99%, it can be judged that the PDS correction is failed. At this condition, the prediction error is often more than 5%. Therefore, the correlation coefficient between the corrected spectrum and its corresponding average spectrum of the homology sample set can be used as an indicator to evaluate the efficiency of the PDS correction.
KeywordsPiecewise direct standardization algorithm; Near infrared quantitative model; Prediction results; Error analysis