-
实验测得的3种样品的3维荧光光谱如图 2所示。由图中可以看出,在3维荧光光谱中,3种样品的最佳激发波长分别为274nm, 285nm, 284nm;荧光峰分别位于450nm, 482nm, 452nm。由图 1可知,3种抗生素的分子结构中均具有苯环和氮杂环构成的共轭环状结构,此外还具有羟基、羧基等荧光助色基团,因此,这3种抗生素都属于荧光物质。3种抗生素在分子结构上的相似性决定了它们具有相似的荧光光谱。但与乳酸环丙沙星相比,另外两种抗生素分子结构中多出一个氮氧杂环,使得整个分子具有更大的共轭结构,因而具有更高的量子产率。而乳酸左氧氟沙星与盐酸左氧氟沙星具有相同的主体分子结构,仅在配体上存在差别。由于乳酸分子更易与主体分子形成氢键,进一步扩大共轭面积,因而相较之下荧光峰值位置也产生一定的红移。
综合考虑3种样品的最佳激发波长,选择285nm的光作为激发光,测量39种浓度的混合溶液样本的常规荧光光谱,测量结果如图 3所示。由图 3可知,随着浓度的改变,荧光峰的位置不发生变化,但荧光强度发生显著改变,荧光强度与浓度之间存在着复杂的非线性关系,难以通过解析函数直接表示。
-
径向基函数神经网络能够逼近任一非线性函数,结构简单、收敛速度快,在不同领域内得到了广泛应用。它分为3层,输入层,隐含层,输出层。不同层有着不同的功能。径向基神经网络分为正则化网络和广义网络。实际应用中,一般都使用广义网络。
广义径向基函数网络的结构图如图 4所示。其中,x1,x2,…,xM为输入层神经元,M为层输入神经元个数;y1, …, yJ为输出层神经元, J为输出层神经元个数; Φ为隐含层节点;Wij表示第i个隐含节点到第j个输出节点的权值; N为训练样本个数。
将配制好的39组混合液样本分为训练组和预测组,取其中35组作为训练组,其余4组作为预测组。每一种浓度的混合液对应一条荧光光谱曲线,每一条曲线有160个波长测量点,将这些测量点对应的荧光强度值归一化后全部作为网络输入值。因此,本实验使用的广义网络输入层有160个节点, 即M=160;输出节点数为3个, 即J=3,分别代表3种抗生素的浓度值。隐含层有I个节点(I<K, K为样本个数),第i个隐含层节点的基函数为:
$ \Phi (\left\| {\mathit{\boldsymbol{X}} - {\mathit{\boldsymbol{X}}_i}} \right\|){\rm{ }} $
(1) 基函数的中心为:
${\mathit{\boldsymbol{X}}_i} = [{x_{i1}}, {x_{i2}}, \ldots , {x_{im}}] $
(2) 隐含层神经元个数的确定采用从零开始递增方法,每增加一个神经元都能最大限度地降低误差,直到满足精度要求。
设实际输出为:
${\mathit{\boldsymbol{Y}}_k} = [{y_{k1}}, {y_{k2}}, \ldots , {y_{kj}}, \ldots , {y_{kJ}}] $
(3) 式中,下标k为输入向量的序号,表示第k个输入向量的输出;j=1, 2, …, J。那么输入训练样本Xk时,网络第j个输出神经元得出的结果为:
$ {\mathit{\boldsymbol{y}}_{jk}} = {W_{0j}} + \sum\limits_{i = 1}^I {{W_{ij}}\mathit{\Phi} ({\mathit{\boldsymbol{X}}_k}, {\mathit{\boldsymbol{X}}_i})} $
(4) -
将训练组样本的常规荧光光谱强度值归一化后作为网络输入值,散布常数取为1.42,训练神经网络,并应用训练好的径向基神经网络,对预测组4个样本各组分的浓度进行预测。预测结果如表 1所示。表中,AC表示实际浓度(actual concentration,AC),单位为ng/mL;PC表示预测浓度(prediction concentration, PC),单位为ng/mL;PE表示预测误差(prediction error, PE)。3种成分预测的平均相对误差分别为4.95%, 9.76%, 12.90%。由此看来,基于常规荧光光谱预测结果误差比较大,并不能满足准确定量预测浓度的要求。
Table 1. Concentration prediction based on conventional fluorescence spectra
No. ciprofloxacin lactate levofloxacin lactate levofloxacin hydrochloride AC/
(ng·mL-1)PC/
(ng·mL-1)PE/% AC/
(ng·mL-1)PC/
(ng·mL-1)PE/% AC/
(ng·mL-1)PC/
(ng·mL-1)PE/% 1 31.7 29.5 6.94 316.7 312.3 1.39 31.7 33.6 5.99 2 18.3 17.9 2.18 100.0 69.7 30.3 26.7 21.9 17.97 3 25.0 27.3 9.20 133.3 139.2 4.42 3.3 4.1 24.24 4 6.7 6.8 1.49 200.0 194.1 2.95 23.3 22.5 3.43 average 4.95 9.76 12.90 -
在混合物分析中,混合物组分的荧光光谱往往相互重叠,不易区分,这时常规的荧光分析方法就会受到限制,而同步荧光分析具有简化光谱、提取有用信息、窄化谱带、减少光谱的重叠和降低散射光的影响等优点,适合多组分混合物的分析。本实验测量3种抗生素的同步荧光光谱(等高线),如图 5所示。
由图 5可知,乳酸环丙沙星在276nm处出现了一个荧光峰,乳酸左氧氟沙星在253nm处出现了一个荧光峰,在284nm处出现了一个荧光峰,而盐酸左氧氟沙星在288nm这个位置出现一个荧光峰。与常规荧光光谱相比,同步荧光光谱对谱线具有明显的窄化作用,凸显出各种样品间的差异性。
恒波长同步扫描过程中,Δλ的选取至关重要,它与光谱信息的获取和质量有很大关系,直接影响到同步荧光光谱的形状、带宽和信号强度。以各组分浓度均为1.67ng/mL的混合液样本为例,分别在Δλ=(130~200)nm(步长为2nm)时测量样本的同步荧光光谱,得到混合溶液在不同Δλ下的同步荧光光谱,如图 6所示。
由图 6可知,在同一样本中,随着Δλ的变化,光谱曲线的形状、峰位、荧光强度均发生显著变化。当Δλ逐渐增大时,荧光峰的位置逐渐向短波方向移动,且荧光强度显著增强,256nm和284nm处的荧光峰逐渐凸显出来,在332nm处还出现一个肩峰。当Δλ增大到194nm时,3个荧光峰差异性最明显。故确定Δλ=194nm为最佳扫描波长差。
在Δλ=194nm时,同一测量条件下,测量39个样本的2维同步荧光光谱,如图 7所示。由图 7可以看出,每个样本的同步荧光光谱,均出现两个荧光峰,分别位于256nm和284nm, 一个肩峰位于332nm。随着浓度的变化,这些峰的位置不发生变化,但荧光强度发生显著变化。
-
将样本的同步荧光光谱强度值归一化后作为网络输入值,取散布常数1.42,训练神经网络,并应用训练好的径向基神经网络,对预测组4个样本各组分的浓度进行预测。预测结果如表 2所示。3种成分预测的平均相对误差分别为3.59%, 3.47%, 3.09%。与常规荧光光谱相比,同步荧光光谱预测结果误差显著减小,准确度更高,体现了同步荧光光谱在混合物分析中的优势。
Table 2. Concentrationprediction based on synchronous fluorescence spectra
No. ciprofloxacin lactate levofloxacin lactate levofloxacin hydrochloride AC/
(ng·mL-1)PC/
(ng·mL-1)PE/% AC/
(ng·mL-1)PC/
(ng·mL-1)PE/% AC/
(ng·mL-1)PC/
(ng·mL-1)PE/% 1 31.7 30.7 3.15 316.7 290.5 8.27 31.7 29.5 6.94 2 18.3 18.2 0.54 100.0 99.2 0.80 26.7 27.0 1.12 3 25.0 27.3 9.20 133.3 130.1 2.40 3.3 3.4 3.03 4 6.7 6.6 1.49 200.0 195.2 2.40 23.3 23.6 1.29 average 3.59 3.47 3.09
同步荧光结合神经网络同时测定3种抗生素
Simultaneous determination of three antibiotics based on synchronous fluorescence combined with neural network
-
摘要: 为了更快速、简便、准确地同时测定多种抗生素混合物,采用同步荧光光谱结合径向基神经网络的方法,对3种氟喹诺酮类抗生素(乳酸环丙沙星、乳酸左氧氟沙星、盐酸左氧氟沙星)的同步荧光光谱进行研究。选择3组分浓度均为1.67ng/mL的混合溶液,测量其3维同步荧光光谱;分别测量39种不同浓度的混合溶液样本的同步荧光光谱;选取其中35种作为训练组,其余4种作为预测组,将训练组样本对应的光谱数据作为输入,建立和训练径向基神经网络;在发射波长与激发波长的差Δλ=194nm条件下,利用训练好的神经网络对预测组中各组分的浓度进行预测,得到3种组分浓度预测的平均相对误差分别达到3.59%,3.47%,3.09%。结果表明,当Δλ设定为194nm时,3种抗生素的同步荧光峰差异最为明显、区分度高,该方法能实现对3种抗生素混合物中各组分的同时测定。这为多种抗生素混合物同时测定提供了一种快速、简便、准确的方法。Abstract: In order to determine the antibiotic mixture more quickly, conveniently and accurately at the same time, synchronous fluorescence spectra of 3 kinds of fluoroquinolones (ciprofloxacin, levofloxacin lactate, levofloxacin hydrochloride) were studied based on synchronous fluorescence spectroscopy combined with radial basis function neural network. The 3-D synchronous fluorescence spectrum for the 3-component mixed solution with concentration of 1.67ng/mL was measured. Then, simultaneous fluorescence spectra of 39 mixed solutions with different concentrations were measured. 35 of them were selected as the training group, and the other 4 were used as the prediction group. The spectral data corresponding to the training samples were taken as input to build and train the radial basis function neural network. The results show that, when Δλ=194nm, the concentration of each component in the prediction group is predicted by the trained neural network. The average relative errors of intensity prediction of 3 components were 3.59%, 3.47% and 3.09%, respectively. The method provides rapid, simple and accurate method for simultaneous determination of multiple antibiotic mixtures.
-
Table 1. Concentration prediction based on conventional fluorescence spectra
No. ciprofloxacin lactate levofloxacin lactate levofloxacin hydrochloride AC/
(ng·mL-1)PC/
(ng·mL-1)PE/% AC/
(ng·mL-1)PC/
(ng·mL-1)PE/% AC/
(ng·mL-1)PC/
(ng·mL-1)PE/% 1 31.7 29.5 6.94 316.7 312.3 1.39 31.7 33.6 5.99 2 18.3 17.9 2.18 100.0 69.7 30.3 26.7 21.9 17.97 3 25.0 27.3 9.20 133.3 139.2 4.42 3.3 4.1 24.24 4 6.7 6.8 1.49 200.0 194.1 2.95 23.3 22.5 3.43 average 4.95 9.76 12.90 Table 2. Concentrationprediction based on synchronous fluorescence spectra
No. ciprofloxacin lactate levofloxacin lactate levofloxacin hydrochloride AC/
(ng·mL-1)PC/
(ng·mL-1)PE/% AC/
(ng·mL-1)PC/
(ng·mL-1)PE/% AC/
(ng·mL-1)PC/
(ng·mL-1)PE/% 1 31.7 30.7 3.15 316.7 290.5 8.27 31.7 29.5 6.94 2 18.3 18.2 0.54 100.0 99.2 0.80 26.7 27.0 1.12 3 25.0 27.3 9.20 133.3 130.1 2.40 3.3 3.4 3.03 4 6.7 6.6 1.49 200.0 195.2 2.40 23.3 23.6 1.29 average 3.59 3.47 3.09 -
[1] HUANG Y S, LIU B P, ZHU X L, et al. A rapid high-performance liquid chromatography-tandem mass spectrometric method for determination of 4 fluoroquinolones residues in fish meat[J].Food Science, 2010, 31(2): 127-130(in Chinese). [2] LI Z M, SUN Y N, YAO K A, et al. Progress of pretreatment and analytical methods for antibiotic residues in animal derived food[J]. Chinese Journal of Pharmaceutical Analysis, 2013, 33(6): 901-906(in Chinese). [3] WU L L, BI Y P, LI D, et al. Detection of antibiotic residues in food by microbiological technology[J]. Chinese Journal of Microecology, 2014, 26(5):598-599(in Chinese). [4] SRINIVASAN P, SARMAH A K, MANLEY-HARRIS M, et al. Development of an HPLC method to analyze four veterinary antibiotics in soils and aqueous media and validation through fate studies[J]. Journal of Environmental Science and Health, 2012, 47(13):2120-2132. [5] GANNA F, NEBESKY V, RANDAK T, et al. Simultaneous determination of 32 antibiotics in aquaculture products using LC-MSMS[J]. Chemical Papers-Slovak Academy of Sciences, 2014, 68(1):29-36. [6] CHEN G Q, WU Y M, LIU H J, et al. Determination and identification of synthetic food colors based onfluorescence spectroscopy and radial basis function neural networks[J].Spectroscopy and Spectral Analysis, 2010, 30(3):706-709(in Chinese). [7] GU L, HUANG Q F, MA C Q, et al. Characteristics and composition analysis of pure liquid milk based on 3-D fluorescent spectra[J]. Laser Technology, 2011, 35(6):742-744(in Chinese). [8] CARSTEA E M, BRIDGEMAN J, BAKER A, et al. Fluorescence spectroscopy for wastewater monitoring A review[J].Water Research, 2016, 95(10):205-219. [9] SÁDECKÁ J, JAKUBÍKOVÁ M, MÁJEK P, et al.Classification of plum spirit drinks by synchronous fluorescence spectroscopy[J]. Food Chemistry, 2016, 196(7):783-790. [10] ZHANG Y, CHENG Y, YAN Y T, et al. Study on characteristics of fluorescent spectra of several kinds of heated common edible oil[J]. Laser Technology, 2013, 37(1):109-113(in Chinese). [11] XU J G, WANG Z B. Fluorescence analysis[M]. Beijing: Science Press, 2006:137-140(in Chinese). [12] ZHAO J H, YUAN H C, XIAO H B, et al. Determination of gentamicin content in duck egg white by using derivative synchronous fluorescence spectrum and GA[J]. Journal of Optoelectronics·Laser, 2014, 25(4): 735-740(in Chinese). [13] XIAO H B, ZHAO J H, YUAN H C, et al. Determination of tetracycline content in pork by synchronous fluorescence with CARS method[J].Optics and Precision Engineering, 2013, 21(10): 2513-2519(in Chinese). doi: 10.3788/OPE. [14] LIU Y Zh, YANG C H, ZHU Y X, et al. Simutaneous determination of dissolved phenanthrene and 2-ethylphenenthrene in aqueous solution by synchronous fluorescence[J]. Journal of Instrumental Analysis, 2015, 34(12): 1366-1371(in Chinese). [15] GE Z X, SUN Z Q. Neural network theory and achievement by MATLAB R2007[M]. Beijing: Publishing House of Electronic Industry, 2007: 4-8, 108-123(in Chinese).