【作者】 梁杰；

【导师】 崔洪亮；

【作者基本信息】 吉林大学 ， 测试计量技术及仪器， 2021， 博士

【摘要】 结核病是全球十大高感染、高致死率传染病之一,并且我国是仅次于印度居于世界第二的结核病高负担国家。因此,我国俨然将治疗结核病列为亟需关注的公共卫生问题之一,且在《“十三五”全国结核病防治规划》中把确保抗结核药品质量列为治疗结核病的首要任务。目前,治疗结核病推荐用药是抗结核复方药物。复方药物的药效是建立在多种主要成分综合作用基础之上;若有效成分含量不足,不但大大降低治愈功效,甚至可能因抗药性肺结核菌株的生长导致病人产生耐药性,从而促使病情进一步加重,所以抗结核复方药物中各有效成分含量的准确检测是其质量控制的关键环节。目前常用的抗结核复方药物成分含量检测方法主要存在检测步骤繁琐、检测时间长、检测试剂污染环境等问题,因此亟需寻求一种简单、快速、绿色的抗结核复方药物含量检测方法。太赫兹时域光谱(THz-TDS)技术作为一种新兴光谱检测技术,因对被测物质几乎无电离损伤且对药物分子集体运动模式(晶格振动、同分异构体构型差异等)非常敏感,在医药质量检测领域备受关注。本文以由异烟肼(INZ)、吡嗪酰胺(PNZ)和利福平(RIF)为有效成分组成的二联抗结核复方药物(INZ+RIF)和三联抗结核复方药物(INZ+RIF+PNZ)为研究对象,以THz-TDS为主要检测技术,展开抗结核复方药物中有效成分含量检测的深入研究。首先理论分析并仿真解析了有效成分INZ、RIF和PNZ的太赫兹光谱,为复方药物中有效成分的定性鉴别和定量分析提供理论依据;接着分别利用集成学习算法和卷积神经网络结合THz-TDS技术对二联和三联复方药物中各成分含量检测展开应用研究。具体的研究内容与研究成果如下:(1)INZ、RIF和PNZ的太赫兹光谱解析研究。利用THz-TDS技术获取了INZ、PNZ、RIF的太赫兹吸收光谱,并利用密度泛函理论(DFT)进行了基于单分子、二聚体以及晶体结构的理论仿真。实验结果显示:INZ在0.3-1.8THz波段内存在3个振动吸收峰,分别位于1.16THz、1.46THz以及1.56THz,其中位于1.16THz处的振动吸收峰来自于分子内相互作用,而位于1.46THz和1.56THz来自于分子间相互作用;PNZ在0.3-1.8THz波段内存在3个振动吸收峰,分别位于0.50THz、0.71THz以及1.42THz,这3个特征吸收峰均自于分子间相互作用;RIF通过实验和仿真计算均没有得到明显的特征吸收峰。该研究结果表明,这3种抗结核复方药物主要成分在太赫兹波段具有明显不同的太赫兹吸收特性,这些不同的光谱吸收特性可作为抗结核复方药物成分的定性鉴别提供一定的理论依据,也为后续定量研究提供了理论支持。(2)为了提高模型的预测准确度和泛化能力,提出了一种基于集成算法结合THz-TDS技术定量建模方法。以梯度提升决策树回归(GBDT)、高斯核支持向量机回归(RBF-SVR)和线性核支持向量机回归(Linear-SVR)为弱学习器,以Linear-SVR为元模型,分别利用Stacking、Blending以及Averaging三种不同的融合算法结合THz-TDS技术建立了二联和三联抗结核复方药物的定量分析模型。研究结果表明:相较于单一模型GBDT、RBF-SVR和Linear-SVR的预测结果,集成模型算法都显著提高了预测性能;其中Stacking集成模型的预测准确度和模型泛化能力最佳,其对所有分析物预测得到的R~2均大于0.99;对于二联抗结核复方药物中INZ和RIF的预测得到的均方根误差RMSEP分别为0.686%和1.009%,对于三联抗结核复方药物中INZ、RIF和PNZ预测得到的RMSEP分别为0.717%、1.021%,0.607%。(3)联合应用CNN和THz-TDS技术建立抗结核复方药物的定量分析模型。首先针对CNN模型的整体架构、重要结构参数、分类器选择、损失函数定义和训练方法进行较为全面的优化与改进,确定最佳CNN架构并完成对二联、三联抗结核复方药物中各组成成分的定量分析。实验结果显示:相较于Stacking模型,CNN模型进一步提高了预测准确度,其针对所有分析物预测得到的RMSEP均小于0.8%。然后,为了探索CNN模型中卷积核对光谱的处理机制,对CNN模型进行了可视化分析,可视化结果分析显示CNN能够完成对光谱不同的转换,提取光谱不同层次的特征,进一步证明了CNN在太赫兹光谱定量分析中具有独特的优势。更多还原

【Abstract】 Tuberculosis is one of the top ten infectious diseases with high infection and high fatality rate in the world,and our country is the second high-burden country of tuberculosis in the world after India.In our country,the treatment of tuberculosis has become one of the public health issues that need urgent attention.In the"Thirteenth Five-Year"National Tuberculosis Prevention and Control Plan,it clearly states that the primary task of treating tuberculosis is to ensure the quality of anti-tuberculosis drugs.Currently,the recommended drug for the treatment of tuberculosis is an anti-tuberculosis compound drug.The efficacy of compound drugs is based on the comprehensive action of multiple main ingredients.If the content of active ingredients is insufficient,not only can not cure the patient,but also lead to the growth of drug-resistant tuberculosis strains,which will cause the patient to develop drug resistance and further aggravate the disease.Therefore,the detection of the content of each active ingredient in the anti-tuberculosis compound medicine is an important aspect of its quality control.At present,the commonly used methods for detecting the content of anti-tuberculosis compound drugs mainly have problems such as cumbersome testing steps,long testing time,and environmental pollution by the testing reagents,etc.Therefore,it is urgent to find a simple,fast and green method for detecting the content of anti-tuberculosis compound drugs.Terahertz time-domain spectroscopy(THz-TDS)technology is an emerging spectroscopic detection technology.Because it generally does not cause ionization damage to substances,and is very sensitive to the collective motion mode of drug molecules(lattice vibration,isomer configuration differences,etc.),it is widely used in the field of medical testing.In this paper,the dual anti-tuberculosis compound drug(INZ+RIF)and triple anti-tuberculosis compound drug(INZ+RIF+PNZ)composed of isoniazid(INZ),pyrazinamide(PNZ)and rifampicin(RIF)as the research object were studied,and took the THz-TDS as the main technical line.Firstly,the terahertz spectra of INZ,PNZ and RIF were analyzed,and then the integrated learning algorithm and convolutional neural network combined with THZ-TDS technology were used to carry out the application research on the content detection of each component in the dual and triple compound drugs,respectively.The specific research contents and results are as follows:(1)Researched the terahertz spectra of INZ,RIF and PNZ.The terahertz absorption spectra of INZ,PNZ,and RIF were obtained using THz-TDS technology,and theoretical simulations based on single molecules,dimers and crystal structures were performed using density functional theory(DFT),respectively.The experimental results showed that:INZ had three vibration absorption peaks,located at 1.16THz,1.46THz and 1.56THz respectively,in the band of 0.3-1.8THz.The vibration absorption peak at 1.16THz came from intramolecular interactions,while the vibration absorption peak at 1.46THz and 1.56THz came from intermolecular interactions;PNZ had three vibration absorption peaks,located at 0.50THz、0.71THz and 1.42THz respectively,in the band of 0.3-1.8THz,which all came from intermolecular interactions.RIF had not obtained obvious characteristic absorption peaks through experiments and simulation calculations.The results of this study show that the main components of the three anti-tuberculosis compound drugs have significantly different terahertz absorption characteristics in the terahertz band,which can provide theoretical basis for qualitative identification and quantitative analysis of active components in compound drugs.(2)In order to improve the prediction accuracy and generalization ability of the model,a quantitative modeling method based on integrated algorithm combined with THz-TDS technology was proposed.Three different fusion algorithms,stacking,blending,and averaging,combined with THz-TDS technology,were used to establish a quantitative analysis model of dual and triple anti-tuberculosis compound drugs.In these fusion algorithms,gradient boosting decision tree regression(GBDT),Gaussian kernel support vector machine regression(RBF-SVR)and linear kernel support vector machine regression(Linear-SVR)are used as elementary learner,and the linear-SVR is used as secondary learner.The experimental results show that:Compared with the prediction results of a single model GBDT,RBF-SVR and Linear-SVR,all the three integrated model algorithms improved the prediction performance significantly,among which the stacking integrated model had the best prediction accuracy and model generalization ability.The R~2 predicted by the stacking integrated model for all analytes was greater than 0.99.The root mean square error(RMSEP)of stacking model for the prediction of INZ and RIF in dual anti-tuberculosis compound drug was 0.686%and1.009%,respectively,and the root mean square error(RMSEP)of stacking model for the prediction of INZ、PNZ and RIF in triple anti-tuberculosis compound drug was0.717%、1.021%,0.607%,respectively.(3)The quantitative analysis model of anti-tuberculosis compound drugs was established by CNN combined with THZ-TDS technology.The CNN was an"end-to-end"quantitative modeling method.Due to its unique internal convolutional kernel structure,CNN model had a strong ability to extract spectral features.First of all,the overall architecture of the CNN model,important structural parameters,classifier selection,loss function definition and training methods were optimized and improved.Then,the best CNN architecture was determined,which was used to complete the quantitative analysis of the components in the dual and triple anti-tuberculosis compound drugs.The experimental results showed that:Compared with the stacking model,the CNN model further improved the prediction accuracy.The RMSEP predicted by the CNN model for all analytes was less than 0.8%.Finally,in order to explore the processing mechanism of the convolution check spectrum in the CNN model,the CNN model was analyzed visually.The analysis of the visualization results showed that CNN could complete the conversion of different spectra and extract the characteristics of different levels of spectra,which further proved that CNN had unique advantages in the quantitative analysis of terahertz spectra.更多还原