药物毒性预测方法研究

发布时间：2018-05-16 19:00

本文选题：药物毒性预测 + 定量构效关系　；参考：《浙江大学》2014年博士论文

【摘要】：新药研发失败约有30%是由于存在安全性问题而导致的。因此,研究建立高效准确的药物毒性预测方法,对于提高新药研发的成功率具有重要意义,并己成为当今毒理学、药物分析学、计算化学和系统生物学等多个学科共同关注的前沿命题。传统的毒理学实验方法由于存在周期长、花费高以及需要消耗大量动物等缺陷,正日益被基于化合物结构的毒性预测方法(如定量构效关系,QSAR)和基于系统生物学的毒性预测方法(如毒理基因组学)所替代。其中,QSAR方法不依赖于实验,具有快速、经济等优点,已被广泛用于药物研发初期进行化合物的毒性评价,但这类方法对化学结构多样且毒性位点或致毒机理差别较大的化合物预测效果较一般；毒理基因组学方法的适用范围较广,所建的基于生物标记物的毒性预测模型有助于加深人们对致毒机理的理解,其缺点是实验费用昂贵,且基于基因芯片的建模预测技术的可靠性尚存争议。鉴于当前毒性预测方法准确率普遍不高的现状,并针对不同的毒性问题,本论文研究建立和改良了多种基于QSAR和毒理基因组学的毒性预测方法： 1、研究提出了一种可用于药物毒性预测的改进决策森林算法(IDF)。通过使用2个高维的基因组数据集进行测试并与多种常用的预测方法进行比较,结果表明,IDF算法不仅获得了比原始决策森林算法更优的结果,而且在准确率和稳定性上也优于其他多种常用的预测方法,说明把IDF算法用于对高维数据进行建模预测具有一定的优势。 2、研究提出了基于SVM(支持向量机)、kNN(k最近邻法)和NC (Nearest Centroids,最近质心)等算法的集成毒性预测方法。基于多个数据集的测试结果表明,集成算法能够大大提高预测准确率,而且结果更稳定,其中集成SVM的方法最优,预测准确率提升超过3%。 3、通过使用来自大鼠血液的基因组表达数据建模,研究提出了可以用于预测药物肝毒性的跨组织预测方法。基于一批来自大鼠血液的基因组表达数据,本研究建立了多个可用于预测药物肝毒性的模型,并在3个独立的肝毒性数据集上进行了验证,其中最高预测准确率达92%。此外还发现了6个血液基因可以作为药物肝毒性的生物标记物。 4、研究提出了基于SVM和微粒群算法的P-糖蛋白底物预测方法。预测一个化合物是否P-糖蛋白底物,对研究该化合物的ADME/T性质具有重要意义。因此,本研究建立了一种基于SVM和微粒群算法的P-糖蛋白底物预测方法。与已有文献结果相比,本方法获得了更高的预测准确率(约90%),且所建模型有更好的化学或生物学意义,可解释性强。
[Abstract]:About 30% of new drug R & D failures are caused by the existence of safety problems. Therefore, it is of great significance to study the efficient and accurate method of predicting drug toxicity, and is of great importance to the success rate of new drug research and development, and has become the frontier life of many subjects such as toxicology, drug analysis, computational chemistry and system biology. Question.
Traditional toxicology experimental methods are being replaced by toxicity prediction methods based on compound structure (such as quantitative structure-activity relationship, QSAR) and toxicity prediction based on System Biology (such as toxicological group). The QSAR method is not dependent on the experiment, because of the long period, high cost and the need to consume a large number of animals. With the advantages of rapid, economic, and so on, it has been widely used in the initial stage of drug development to evaluate the toxicity of compounds. However, this method has a more general prediction effect on chemical compounds with a wide variety of toxic sites or toxic mechanisms, and a wide range of application of toxicological genomics methods. The measurement model helps to deepen people's understanding of the toxic mechanism. The disadvantage is that the experimental cost is expensive, and the reliability of the modeling prediction technology based on the gene chip is still in dispute.
In view of the prevalence of current toxicity prediction methods, and in response to different toxicity problems, a variety of methods of toxicity prediction based on QSAR and toxicology genomics have been established and improved in this paper.
1, the study proposed an improved decision forest algorithm (IDF), which can be used to predict drug toxicity. By using 2 high dimensional genomic data sets to test and compare with a variety of common prediction methods, the results show that the IDF algorithm not only obtained better results than the original decision forest algorithm, but also in accuracy and stability. It is also superior to many other commonly used prediction methods, which shows that IDF algorithm has certain advantages in modeling and forecasting high-dimensional data.
2, the study proposes an integrated toxicity prediction method based on SVM (support vector machine), kNN (k nearest neighbor) and NC (Nearest Centroids, the nearest centroid). The test results based on multiple data sets show that the integrated algorithm can greatly improve the accuracy of prediction, and the result is more stable, in which the method of integrating SVM is optimal and the prediction accuracy is proposed. Rise more than 3%.
3, by modeling the genome expression data from the rat's blood, the study proposed a cross tissue prediction method that could be used to predict the toxicity of the drug. Based on the genome expression data from a group of rat blood, a number of models that could be used to predict the toxicity of drug hepatotoxicity were established and advanced in 3 independent hepatotoxicity data sets. The highest prediction accuracy was 92%., and 6 blood genes were identified as biomarkers of drug hepatotoxicity.
4, the study proposed a method of P- glycoprotein substrate prediction based on SVM and particle swarm optimization. Predicting whether a compound is P- glycoprotein substrate is of great significance to the study of the ADME/T properties of the compound. Therefore, a method for predicting the substrate of P- glycoprotein based on SVM and particle swarm optimization is established. The method has a higher prediction accuracy (about 90%), and the model has better chemical or biological meaning and is more explanatory.

【学位授予单位】：浙江大学
【学位级别】：博士
【学位授予年份】：2014
【分类号】：R994.3

【参考文献】