【作者】 田盛；

【导师】 侯廷军；

【作者基本信息】 苏州大学 ， 化学生物学， 2014， 博士

【摘要】 近年来，随着越来越多的天然产物成功地通过FDA认证而上市，中药（Traditional Chinese Medicines，TCMs）作为天然产物的重要组成部分，在现代药物研发中受到了越来越多的关注和重视。中药用于治疗疾病的主要形式是通过含有多种中草药植物的中药复方来实现的，因此人们普遍认为，中草药可以作为药物研发很好的类药化合物来源。从传统中草药中寻找到相关靶点的潜在活性化合物并确定其药理活性已经成为制药公司药物开发的一个重要途径。人们对基于中草药资源的药物研发已经做过了大量的尝试和研究，但我们对中草药化合物的分子的性质、结构以及成药性特征还缺乏深入的了解。此外，相比较于西药治病理论，大部分中草药治疗疾病的机制都还不够清晰，能否从分子水平阐述中草药治疗相关疾病的作用机制是非常重要的研究课题。最后，如何从中草药化合物中筛选得到相关靶点的潜在活性化合物也是一个热点研究方向。本论文系统开展了基于中草药有效成分的计算机辅助药物分子设计研究。首先，我们系统比较了药物数据库MDDR、非药数据库ACD和中草药化合物数据库（TCMCD）中化合物的物理化学性质以及结构特征的差异。结果表明，相比MDDR和ACD，TCMCD中的化合物性质分布更为广泛并且结构更为复杂和新颖。同时，我们发现基于简单性质的类药性预测规则预测能力较差。为了对中草药化合物的类药性进行定量评价，我们用机器学习方法，包括朴素贝叶斯和递归分割方法，构建了精确的类药性定量预测模型。结果表明，基于分子理化性质描述符构建的类药性模型的预测精度较低，而引入了分子指纹描述符后，类药性模型的预测精度有了较大的提升。同时，我们发现类药性模型的预测能力与训练集的大小以及构成有着直接的关系，用所构建的最为可靠的类药性模型对中草药化合物数据库进行了类药性的评价，超过60%的中草药化合物被预测为类药，表明TCMCD从统计上讲是类药的，可以作为药物研发的一个很好的类药化合物来源。中药治疗疾病主要是通过由多种中草药植物所组成的中药复方的形式发挥作用，因此，由大量中药有效成分构成的中药复方的治疗疾病的机制很不清晰。为了从分子水平阐述中草药复方治疗疾病的机制，我们以治疗Ⅱ型糖尿病中药复方为例进行研究。首先，收集已知治疗Ⅱ型糖尿病的中药复方中含有的有效成分化合物以及与Ⅱ型糖尿病相关的靶点。随后采用分子对接、药效团映射以及机器学习的方法筛选出各靶点的潜在活性化合物。通过构建潜在活性化合物和靶点的相互作用网络，从一定程度上揭示了中草药复方治疗Ⅱ型糖尿病的机制：中药复方中的大部分有效成分只能跟单一靶点形成相互作用，构成治疗Ⅱ型糖尿病的主要作用力，其次，中药复方中的少部分化合物能和多个Ⅱ型糖尿病相关靶点作用，发挥治疗糖尿病的次要作用，协同增强治疗糖尿病的效果，最后，中草药中的部分化合物不与Ⅱ型糖尿病相关靶点形成直接的相互作用，而是通过其他的一些药理活性，如去自由基功能/抗氧化能力、抗菌能力来协助治疗糖尿病及其并发症。所得到的这些结论能够较好的与经典中医药治病理论“君臣佐使”相吻合。为了从中草药化合物数据库TCMCD中筛选得到相关靶点理想的潜在活性化合物，我们以激酶靶点ROCK1为例展开研究。考虑到蛋白柔性对虚拟筛选结果的影响，我们用机器学习方法整合ROCK1靶点多个复合物结构所得到的分子对接和药效团模型的预测结果，构建了新颖的并行虚拟筛选策略并对其预测能力进行了评测。研究结果表明，相比较于基于单个复合物结构的分子对接或药效团模型的预测结果，整合的虚拟筛选策略更为可靠。随后，用构建的并行虚拟筛选策略对中草药化合物数据库进行了虚拟筛选，得到了53个结构新颖的ROCK1潜在活性化合物。这些化合物可以作为理想的ROCK1潜在活性化合物来进行后续的研究。所构建的并行虚拟筛选策略也可以作为一个可靠的工具用于药物筛选。更多还原

【Abstract】 In recent years, many drugs approved by the Food and Drug Administration (FDA)directly come from natural products. As an important source of natural products,traditional Chinese medicines (TCMs) are gaining more and more attention in moderndrug discovery pipelines. The classic TCMs are primarily based on a large number ofherbal formulae that are used for the treatment of a wide variety of diseases. It isbelieved that TCMs are a good source of drug-like compounds. Discovery of newbioactive compounds from herbs used in TCMs and identification of theirpharmacological effects are becoming a promising way for finding new drugs in thepharmaceutical industry. However, until now the in-depth analyses of compoundsidentified in TCMs are still lacking. For example, we do not have in-depthunderstanding about the characteristics of the physicochemical properties, structures anddrug-likeness of the compounds in TCMs. Besides, compared with theory of westernmedicine treatment, the mechanism of TCMs for curing disease is not clear. Uncoveringthe underlying action mechanisms of TCMs for combating diseases at the molecularlevel is an important topic. At last, how to identify promising active compounds moreeffective for targets of interest is also a research hotspot.In order to promote the development and modernization of TCMs, the systematicalstudies on the computer-aided drug design (CADD) based on the compounds in TMCswere reported in our thesis. First, the molecular properties and structural features amongthe drug-like compounds in MDDR, the non-drug-like compounds in ACD and thenatural compounds in TCMCD were investigated systematically. The resultsdemonstrated that, compared with the compounds in MDDR and ACD, the naturalcompounds in TCMCD had more diverse property distribution, novel and morecomplex structural features. In addition, the drug-likeness filters based on simplemolecular properties and/or structural features are unreliable and have low predictionaccuracy. In order to construct more reliable theoretical models for drug-likeness andevaluate the drug-likeness of TCMCD, machine learning techniques, including na ve Bayesian classification and recursive partitioning methods were used. The drug-likenessmodels based on molecular physicochemical properties cannot give satisfactoryprediction accuracy. By adding molecular fingerprints, the prediction power can beimproved substantially. Besides, it can be found that the prediction accuracy of thedrug-likeness model is closely related to the size and the balance degree of the trainingset. Then, the best drug-likeness model to employed to evaluate the drug-likeness of thecompounds in TCMCD and found that more than60%compounds were predicted to bedrug-like. The results indicated that the TCMCD is drug-like statistically and believedto be a good source of drug-like compounds.It is well known that basic form of TCMs for curing diseases is TCM formulae (orprescriptions), which is a mixture of special herbs. Therefore, it is not clear that how alarge number of chemical compounds of TCM formulae combat diseases. In order tounderstand the interaction mechanism of TCM formulae at the molecular level, weinvestigated the theory of TCM formulae for treating type2diabetes (T2DM). First, wecollected the T2DM related targets and the chemical compounds in TCM formulae fortreating T2DM. By employing structure-based virtual screening approaches includingmolecular docking, pharmacophore mapping, and machine learning approaches toidentify potential active compounds for T2DM targets. Then, we built the interactionnetwork between the potential active compounds and T2DM related targets. Byanalyzing the compound-target network, we can conclude the mechanism of TCMformulae for curing T2DM as follows: most chemical compounds in TCM formulae canonly interact with an individual target, forming the leading fighting force to combatT2DM. Then, those potential multi-target compounds may influence the T2DM-relatedtargets, forming additional forces to enhance the therapeutic effects. At last, a portion ofthe compounds are responsible for remedying the other related symptoms that areproved to be related to T2DM, such as free radical scavenging/antioxidant andantibacterial activities. All of these observations can be seen as a proper way to revealthe classical theory “Monarch, Minster, Assistant, and Guide” in TCM prescriptions atthe molecular level.In order to identify potential active compounds from TCMCD for targets ofinterest and considering the influence of protein flexibility in virtual screening, we havedesigned and evaluated a parallel virtual screening protocol by integrating the predictionresults from molecular docking and complex-based pharmacophore searching based onmultiple protein structures of ROCK1. It is encouraging to find that the integrated classifiers illustrate much better performance than molecular docking or complex-basedpharmacophore searching based on any single ROCK1structure. Then, the most reliableclassifier was utilized to identify potential inhibitors of ROCK1from TCMs. Thepotential active compounds are novel compared with the known ROCK1inhibitors, andthey can be served as promising starting points for the development of ROCK1inhibitors. The novel parallel VS strategy developed here is quite reliable and can beused as a powerful tool in drug screening.更多还原