基于深度学习的药物活性研究

发布时间：2018-04-05 19:20

本文选题：药物活性　切入点：计算机辅助算法　出处：《新疆大学》2017年硕士论文

【摘要】：随着世界经济的高速发展,基于生物活性分子的药效研究也不断进步。目前,采用组合药理分析、高通量筛选技术、数理统计分析等多种技术,为药物活性的内在属性研究开启了新的征程。但是,由于药物分子自身高纬度、高复杂度的特性,将各类技术转化为实际操作依旧十分艰难,成为科研过程中的瓶颈。因此,从海量的药理数据中,及时有效的确定活性分子,成为科研工作者迫在眉睫的任务。为了迅速找到该问题的解决方法,科研人员选用计算机作为药物发现的辅助工具,较大程度上,计算机的使用为科研工作者的工作进程起到了一定的推动作用。然而,针对药物活性分子研究,现有的大多数科研工作者仅选用某一类计算方法对其进行探讨,这种研究方式在一定程度上约束了药物活性分子的检测范围,对于药物的及时发现也不利。在实际检测过程中,无标签数据较有标签数据容易获得。因此,根据样本属性,本文采用两类计算机辅助算法(浅层机器学习和深层机器学习)进行药物活性分子研究,浅层机器学习选用有监督算法(Supervised algorithm)和半监督算法(Semi-supervised algorithm),深层机器学习选用无监督算法(Unsupervised algorithm)。有监督算法中,支持向量机(Support Vector Machine,SVM)和人工神经网络(Artificial Neural Network,ANN)较为常见。半监督算法中,半监督支持向量机(Semi-supervised support vector machine,S4VM)和代价安全性半监督支持向量机(Cost security semi-supervised support vector machine,CS4VM)较有代表性。无监督算法中,栈式自编码(Stacked AutoEncoder,SAE)和深度信念网络(Deep Belief Network,DBN)较为杰出。针对研究目的,本文将此六种方法进行合理分配,分别对三类药物活性分子(PLK1PBD、SMAD3、IL-1B)进行深入探究。由于药物活性分子结构繁杂,选用化学计量学软件MOE对其进行精密计算,分别获得其2D及3D分子描述符,通过上述两类算法进行药物活性分子识别,实验结果表明,在相同条件下,基于深度学习的无监督算法更容易提取活性分子的深层信息,较其它算法而言,其精确度、敏感度、特异度及命中率等指标皆具有明显优势。
[Abstract]:With the rapid development of the world economy, the research on bioactive molecules has been improved.At present, combinatorial pharmacological analysis, high-throughput screening technology, mathematical statistical analysis and other techniques have opened up a new journey for the study of intrinsic properties of drug activity.However, due to the characteristics of high latitude and high complexity of drug molecules, it is still very difficult to convert all kinds of technologies into practical operation, and become the bottleneck in the research process.Therefore, from the massive pharmacological data, it becomes an urgent task for researchers to identify active molecules in a timely and effective manner.In order to find the solution to the problem quickly, the researchers choose the computer as the auxiliary tool of drug discovery. To a large extent, the use of the computer has played a certain role in promoting the work process of the researchers.However, for the study of drug active molecules, most existing researchers only choose one kind of calculation method to discuss it, this kind of research method restricts the detection range of drug active molecules to some extent.It is also unfavorable to the timely discovery of drugs.In the actual detection process, the untagged data is easier to obtain than the labeled data.Therefore, according to the sample properties, two kinds of computer-aided algorithms (shallow machine learning and deep machine learning) are used to study the active molecules of drugs.The supervised algorithm and semi-supervised algorithm are selected for shallow machine learning and unsupervised algorithm for deep machine learning.Among supervised algorithms, support Vector Machine (SVM) and artificial Neural Network (Ann) are more common.Semi-supervised support vector machine (semi-supervised support vector machine) and cost security semi-supervised support vector machine (CS4VM) are more representative in semi-supervised security semi-supervised support vector machine than semi-supervised support vector machine (semi-supervised support vector machine) and cost security semi-supervised support vector machine (CS4VM).Among the unsupervised algorithms, stackable AutoEncoding (SAE) and Deep Belief Network (DBN) are outstanding.For the purpose of this study, the six methods were divided into three kinds of drug active molecule, PLK1 PBD1, SMAD3, IL-1B, respectively.Because of the complexity of the active molecular structure, the 2D and 3D molecular descriptors are obtained by using the chemometrics software MOE. The experimental results show that the two algorithms are used to recognize the active molecules.Under the same conditions, the unsupervised algorithm based on deep learning is easier to extract the deep information of active molecules, and its accuracy, sensitivity, specificity and hit rate have obvious advantages over other algorithms.
【学位授予单位】：新疆大学
【学位级别】：硕士
【学位授予年份】：2017
【分类号】：R91;TP181

【参考文献】