基于化学键的中药代谢产物预测模型构建方法的建立及其应用研究

发布时间：2018-04-27 05:19

本文选题：代谢产物预测 + 代谢网络预测　；参考：《北京中医药大学》2017年硕士论文

【摘要】：背景:目前,多数中药化学成分体内代谢过程尚不清楚,这在一定程度上阻碍了中药直接作用物质基础和作用机制的阐明。中药化学成分体内代谢产物的研究有助于揭示中药体内变化过程,从而为中药直接作用物质基础和作用机制的阐明提供线索与指导。利用传统实验的方法进行中药化学成分代谢产物研究费时而又耗力,同时高成本、高消耗。而通过计算机模拟构建代谢产物预测模型的方法因其安全、高效、低成本、低消耗等特点有效弥补了传统实验的不足,被认为是辅助药物代谢产物研究的另一有效手段。近年来,越来越多基于计算机模拟的药物代谢预测研究先后被报道,但主要是围绕着代谢位点的预测展开的,少数模型能够给出具体的代谢产物却又无法提供代谢酶的信息,同时模型的可靠性有待提高。因此建立具有较高可靠性的代谢产物预测模型是目前中药代谢预测研究领域亟待解决的问题。目的:本论文旨在构建快速、有效的中药化学成分代谢产物预测模型,弥补当前代谢产物预测模型对于代谢产物和代谢酶二者不可兼得的不足,并将其应用于中药代谢网络的构建。方法:(1)代谢反应数据的考察。从数据的准确性、完备性、适用性和可获得性四个方面对目前已有代谢反应数据进行考察。准确性主要考察数据的原始出处。完备性主要考察数据库对于当前已知代谢反应数据的覆盖度。适用性主要考察数据信息是否完善,考虑到本论文使用的代谢反应数据需催化代谢反应的代谢酶已知,因此对于催化酶未知的代谢反应数据不予考虑。可获得性主要考察数据获得的难易程度。最终,基于准确性高、完备性好、适用性高、易于获得的原则确定本论文的最佳数据源。(2)中药化学成分代谢产物预测模型构建方法。立足于收集的代谢反应数据,根据每个代谢位点发生代谢反应的实际情况,进行阳性代谢位点和未标记代谢位点的划分。利用简单投票的方法进行阴性代谢位点的筛选。将参数化的阳性代谢位点和阴性代谢位点作为输入,基于CHI、IG、GR、Relief四种特征属性选择方法进行特征属性的筛选。立足于筛选所得的特征属性,分别通过 Bayes、LibSVM、KStar、IBK、AdaBoost、Boosting、J48、RandomForest八种分类建模方法进行代谢产物预测模型的构建,结果形成32种不同的建模组合。通过对模型预测性能的比较,获得最优代谢产物预测模型。最终,通过以下三个方面对模型的可靠性进行考察:第一,利用独立测试集,从代谢位点水平上考察模型的可靠性。第二,基于文献报道的经典外部测试集,从化合物分子水平上考察模型的可靠性,并通过与既往文献报道的代谢产物预测模型进行比较,考察模型的预测性能。第三,基于文献报道的中药化学成分代谢数据,考察本论文所构建的代谢产物预测模型在中药化学成分代谢产物预测中的表现。(3)中药化学成分代谢网络预测模型构建方法。基于本论文构建的两套代谢产物预测模型构建中药化学成分代谢网络预测模型,并基于文献报道的中药化学成分代谢网络数据和药物代谢网络数据考察所建中药化学成分代谢网络预测模型的可靠性。结果:(1)CYP450 3A4,2D6和2C9介导代谢反应代谢产物预测模型的构建。基于文献报道的CYP450酶相关代谢反应数据,构建了 CYP450 3A4,2D6和2C9介导的5类代谢反应代谢产物预测模型。涉及的代谢反应类型包括N-脱烷基反应、O-脱烷基反应、脂肪族羟基化反应、芳香族羟基化反应和S-氧化反应。所有模型的10折交叉验证测试的准确性均介于0.940-0.987之间,灵敏度和特异性分别大于0.856和0.968,受试者工作特征曲线下面积均在0.953以上。所有模型独立测试集测试的准确性均介于0.947-0.994之间,灵敏度和特异性分别大于0.864和0.933,受试者工作特征曲线下面积均在0.977以上。对于文献报道的外部测试集,本论文给出的灵敏度和特异性值分别为0.821和0.956,与前人的研究(灵敏度:0.701,特异性:0.963)相比,本论文所建模型识别真阳性样本的能力较前人的研究提高了 12%,识别真阴性样本的能力与前人的研究相比无显著差异。从分子水平上来讲,外部测试集的34个分子中,本论文所建模型完全预测正确17个分子,优于SOMP 13个分子完全预测正确的结果。总的来说,本论文所建模型的预测性能要优于或至少与前人的工作一样好。(2)氧化还原酶介导代谢反应代谢产物预测模型的构建。通过对11个候选代谢反应数据库的考察,最终选定BKM数据库为最佳数据源。基于BKM数据库收录的代谢反应数据,建立了氧化还原酶介导的7类氧化代谢反应代谢产物预测模型。涉及的代谢反应类型包括C=C的形成反应、醇氧化成酮反应、脂肪族羟基化反应、醇氧化成醛反应、芳香族羟基化反应、氧化脱氨反应和N-脱烷基反应,共涉及代谢酶655个。所有模型10折交叉验证的准确性均介于0.901-0.995之间,除醇氧化成酮反应和芳香族羟基化反应的灵敏度分别为0.777和0.765之外,其他5类代谢反应的灵敏度均在0.875-0.988之间,所有模型的特异性和受试者工作特征曲线下面积分别大于0.944和0.915。对于独立测试集而言,所有模型的准确性均介于0.885-0.969之间,除芳香族羟基化反应之外,其他6类代谢反应的灵敏度均大于0.777,所有模型的特异性和受试者工作特征曲线下面积分别大于0.963和0.916。训练集和独立测试集的平衡精度分别大于0.858和0.813。对于文献报道的外部测试集,从代谢位点水平上来讲,34个阳性代谢位点中31个代谢位点完全预测正确,灵敏度为0.912。从分子水平上来讲,外部测试集31个化合物分子中23个分子完全预测正确。结果表明,当涉及的代谢酶更多、数据量更大、数据结构更加复杂时,笔者提出的代谢产物预测模型构建方法依然合理有效。(3)中药化学成分代谢网络预测模型的构建。基于(1)和(2)所建的两套代谢产物预测模型初步建立了中药化学成分代谢网络预测模型。该中药化学成成分代谢网络预测模型的预测范围如下:CYP450 3A4,2D6和2C9介导的N-脱烷基反应、O-脱烷基反应、脂肪族羟基化反应、芳香族羟基化反应,S-氧化反应等5类代谢反应和氧化还原酶介导的C=C的形成反应、醇氧化成酮反应、脂肪族羟基化反应、醇氧化成醛反应、芳香族羟基化反应、氧化脱氨反应、N-脱烷基反应等7类代谢反应。最后,基于文献报道的西酞普兰代谢网络和乌头碱代谢网络考察了所建代谢网络预测模型的可靠性。结果发现,文献报道的西酞普兰代谢过程和乌头碱代谢过程在本论文预测的结果中均有一定程度的体现。结论:本论文首次提出了基于化学键的代谢产物预测模型构建方法。并成功构建了CYP450 3A4,2D6和2C9介导的5类代谢反应代谢产物预测模型和氧化还原酶介导的7类代谢反应代谢产物预测模型。测试结果表明本论文构建的代谢产物预测模型真实、可靠,其预测性能优于或至少与既往文献报道的代谢产物预测模型一样好。最终,在以上两套代谢产物预测模型的基础上,初步构建了中药化学成分代谢网络预测模型。测试结果表明本论文构建的代谢网络预测模型具有一定的可靠性,在一定程度上可以反应中药化学成分的体内代谢过程。总的来说,本论文所建的中药化学成分代谢产物预测模型和中药化学成分代谢网络预测模型将为中药直接作用物质基础的发现和中药作用机制的阐明提供线索,也将为新药开发过程中先导化合物的优化提供策略。
[Abstract]:Background: at present, the metabolic process of most Chinese medicine chemical components is not clear, which hinders the clarification of the material basis and mechanism of the direct action of traditional Chinese medicine. The study of the metabolites in the body of Chinese medicine is helpful to reveal the process of the change in the body of Chinese medicine, and to explain the direct effect of the material basis and mechanism of action for the Chinese traditional medicine. It provides clues and guidance. Using traditional experimental methods to study the metabolites of chemical components of traditional Chinese medicine is time-consuming, consuming, high cost and high consumption, and the method of building the prediction model of metabolites by computer simulation has effectively made up the shortage of traditional experiments because of its safety, efficiency, low cost and low consumption. In recent years, more and more studies have been reported on the prediction of drug metabolism based on computer simulation, but mainly around the prediction of metabolic sites. A few models can give specific metabolites but can not provide the information of metabolic enzymes, and the reliability of the model is also available. Therefore, the establishment of a high reliability metabolic product prediction model is an urgent problem in the field of Chinese medicine metabolism prediction research. Objective: the purpose of this paper is to construct a fast and effective model for predicting the metabolites of the chemical composition of Chinese medicine, and make up for the current metabolic product prediction model for the metabolic products and the metabolic enzyme two. It is applied to the construction of the metabolic network of traditional Chinese medicine. Methods: (1) investigation of metabolic response data. From the four aspects of data accuracy, completeness, applicability and availability, the existing metabolic response data are investigated. The accuracy mainly examines the original source of the data. The coverage of the metabolic response data. Applicability mainly to investigate whether the data information is perfect. Considering that metabolic enzymes used in this paper need to be known as metabolic enzymes in the catalytic metabolic reaction, the metabolic response data of the unknown catalytic enzyme are not considered. High quality, good completeness, high applicability and easy access to the best data sources of this paper. (2) the construction method of the prediction model of the metabolites of chemical components of Chinese medicine. Based on the collected metabolic reaction data, the positive metabolic sites and unlabeled metabolic sites are divided according to the actual situation of metabolic reactions at each metabolic site. The screening of negative metabolic sites was carried out by using a simple voting method. The parameterized positive metabolic sites and negative metabolic sites were used as input to select four characteristic attribute selection methods based on CHI, IG, GR and Relief. Based on the selected characteristics, Bayes, LibSVM, KStar, IBK, AdaBoost, Boosting, respectively, were selected. J48, RandomForest eight classification modeling methods for the construction of metabolic product prediction model, resulting in the formation of 32 different modeling combinations. Through the comparison of the prediction performance of the model, the optimal metabolic product prediction model was obtained. Finally, the reliability of the model was investigated through the following three aspects: first, the use of independent test set, from the metabolic site. The reliability of the model was examined at point level. Second, based on the classical external test set, the reliability of the model was investigated from the molecular level of the compound, and the predictive performance of the model was compared with the previously reported metabolic product prediction model. Third, based on the literature report of the metabolic data of the chemical composition of Chinese medicine, The performance of metabolic product prediction model in this paper is observed in the prediction of metabolic products of Chinese medicine chemical components. (3) construction method of metabolic network prediction model of Chinese medicine chemical composition. Based on the two sets of metabolic product prediction models constructed in this paper, the pretest model of metabolic network of Chinese Medicine chemical composition is constructed, and based on the literature report of Chinese Medicine Chemistry The reliability of the prediction model of the chemical composition metabolic network of Chinese medicine was constructed by the composition metabolic network data and the drug metabolic network data. Results: (1) the construction of the prediction model of metabolic reaction products mediated by CYP450 3A4,2D6 and 2C9. Based on the reported data of the related metabolic reaction of CYP450 enzymes, the 5 types of CYP450 3A4,2D6 and 2C9 are constructed. Metabolic response metabolite prediction models. The types of metabolic reactions involved include N- dealkylation, O- dealkylation, aliphatic hydroxylation, aromatic hydroxylation and S- oxidation. The accuracy of all 90% off cross validation tests of all models is between 0.940-0.987, sensitivity and specificity of more than 0.856 and 0.968, respectively. The area under the working characteristic curve of the subjects is above 0.953. The accuracy of all model independent test sets is between 0.947-0.994, sensitivity and specificity are more than 0.864 and 0.933 respectively. The area under the working characteristic curve of the subjects is more than 0.977. For the external test set reported in the literature, the sensitivity and specificity of this paper are given. The heterosexual values are 0.821 and 0.956 respectively. Compared with previous studies (sensitivity: 0.701, specificity: 0.963), the ability of the model to identify the true positive samples is 12% higher than that of previous studies. The ability to identify the true negative samples is not significantly different from those of previous studies. From the molecular level, the 34 molecules of the external test set are on the molecular level. In this paper, the model is completely predicted by 17 molecules, which is better than SOMP 13 molecules to fully predict the correct results. In general, the prediction performance of the model in this paper is better than or at least as good as the previous work. (2) the construction of the oxidoreductase mediated metabolic product prediction model of the oxidoreductase mediated by 11 candidate metabolites. According to the investigation of the database, the BKM database is selected as the best data source. Based on the metabolic response data of the BKM database, a predictive model of 7 kinds of oxidoreductase mediated metabolic reaction products is established. The metabolic reaction types include the formation reaction of C=C, the alkyl ketone reaction, the aliphatic hydroxylation reaction, and the alcohol oxidation. Aldehyde reaction, aromatic hydroxylation, oxidative deamination and N- dealkylation involved 655 metabolic enzymes. The accuracy of all 90% off cross validation of all models was between 0.901-0.995, and the sensitivity of alcohol oxide to ketone reaction and aromatic hydroxylation was 0.777 and 0.765 respectively, and the sensitivity of the other 5 types of metabolic reactions were all In 0.875-0.988, the specificity of all models and the area under the working characteristic curve of the subjects were greater than 0.944 and 0.915. respectively. For the independent test set, the accuracy of all models was between 0.885-0.969. In addition to the aromatic hydroxylation, the sensitivity of the other 6 types of metabolic reactions was greater than 0.777, and the specificity of all models and the specificity of all models were more than 0.777. The area under the working characteristic curve of the subjects was greater than 0.963 and the 0.916. training set and the independent test set were more than 0.858 and 0.813. for the external test set. From the level of metabolic sites, 31 metabolic sites in 34 positive metabolic sites were completely pretested and the sensitivity was 0.912. from the molecular level. 23 of the 31 compounds in the external test set are correctly predicted. The results show that when the metabolic enzymes involved are more, the amount of data is larger, and the data structure is more complex, the method of building the metabolic product prediction model proposed by the author is still reasonable and effective. (3) the construction of the prediction model of the metabolic network of Chinese medicine. Based on (1) and (2) The two set of metabolic product prediction models initially established the prediction model of the chemical composition metabolic network of traditional Chinese medicine. The prediction range of the chemical composition metabolic network model of the traditional Chinese medicine is as follows: CYP450 3A4,2D6 and 2C9 mediated N- dealkylation, O- dealkylation, aliphatic hydroxylation, aromatic hydroxylation, S- oxidation, etc. 5 kinds of metabolic reactions and oxidoreductase mediated C=C formation reaction, alcohol oxidation to ketone reaction, aliphatic hydroxylation, alcohol oxidation to aldehyde reaction, aromatic hydroxylation, oxidation deamination, N- dealkylation, etc. Finally, based on the literature report of citalopram metabolic network and aconitine metabolic network The results showed that the metabolic processes of citalopram and the metabolic process of aconitine were reflected to some extent in the results of this paper. Conclusion: This paper first proposed a method to construct the prediction model of metabolites based on chemical bonds, and successfully constructed CYP450 3A4,2D6 and 2. The C9 mediated metabolic product prediction model and the oxidoreductase mediated metabolic product prediction model of the 7 metabolic reaction mediated by the oxidoreductase. The results show that the predictive model of the metabolites constructed in this paper is true and reliable, and its predictive performance is as good as or at least as good as that of the previously reported metabolite prediction model. On the basis of two sets of metabolic product prediction models, a preliminary construction of the prediction model of the chemical composition metabolic network of Chinese medicine has been built. The results show that the metabolic network prediction model constructed in this paper is reliable and can reflect the metabolic process in the body of the chemical composition of Chinese medicine to a certain extent. In general, the Chinese medicine chemistry built in this paper The prediction model of composition metabolites and the prediction model of chemical composition metabolic network of Chinese medicine will provide clues for the discovery of the material basis of the direct action of traditional Chinese medicine and the clarifying of the mechanism of action of traditional Chinese medicine, and will also provide a strategy for the optimization of the pilot compounds in the process of the development of new drugs.

【学位授予单位】：北京中医药大学
【学位级别】：硕士
【学位授予年份】：2017
【分类号】：R284

【参考文献】