基于生物网络特征和药物基因组学特征的抗癌协同组合药物预测方法研究

发布时间：2018-05-18 12:51

本文选题：抗癌协同组合药物 + 药物靶点网络特征　；参考：《上海海洋大学》2017年硕士论文

【摘要】：复杂疾病如癌症等往往是由多基因异常引起的系统性疾病,基因突变、拷贝数变异以及基因表达失衡等均可能成为肿瘤发生的原因或表现。传统的“一药一靶”药物研发策略已经被证明存在一定的缺陷,比如容易产生耐药性和毒副作用等。此外生物系统的鲁棒性使得单一药物很难完全抑制肿瘤细胞的生长。肿瘤的化学治疗要求药物可以作用于多个靶点和多条信号通路,并从多个靶点和通路上阻止疾病相关信号的发生、传导和作用。组合药物疗法是多靶点治疗中最常使用的一种疗法,其具有毒性低、效率高等优点。因此建立协同组合药物筛选的方法具有重要意义。筛选协同组合药物的传统方法是实验筛选法,然而实验筛选法往往成本很高且效率低下,只能发现少量具有协同作用的药物组合,而基于计算机模拟的方法对协同药物进行筛选具有通量大、效率高、成本低和污染少等优点。随着多组学数据的大量积累和生物信息学技术的不断发展,使用计算机模拟的方法有望代替实验筛选法进行协同药物筛选。因此对于肿瘤研究人员来说,建立一种有效的预测抗癌协同药物的计算模型将会具有重要意义。近年来,基于计算模型预测协同药物的研究发展迅速。根据构建预测模型时使用机器学习算法的不同,可将这些模型分为无监督学习预测模型、半监督学习预测模型和监督学习预测模型。其中监督学习预测模型中较常使用的是随机森林算法(Random Forest)。基于随机森林算法的模型需要适量的有标记样本做训练集数据,然后进行特征设计,再对有标记样本进行特征值打分。这些预测模型的核心是所设计的特征。这些特征往往是基于药物表型数据设计的,如药物靶点、化学结构等。尽管这些特征也是药物的重要特征,但是这些特征并不能全面地反映药物的药理学特性。药物基因组学数据(药物处理细胞系后的表达谱数据)可以精确地反映药物对细胞系的扰动情况,能够从一定程度上反映药物的药理学特性。因此特征设计时,需要综合考虑药物表型特征和药物基因组学特征。本研究以DREAM Challenge 7 sub-challenge 2中已知的协同药物组合、叠加药物组合和拮抗药物组合为基础,结合相应的药物表型数据和药物处理细胞系后的基因表达谱数据,设计和分析了21个组合药物的特征(包括药物表型特征和药物基因组学特征)。通过使用随机森林算法进行特征筛选、模型构建,得到了最优预测模型。然后我们筛选出Connectivity Map数据库中美国食品与药品监察管理局(U.S.FDA)已批准的抗癌药物。并将这些药物表型数据及其相应的基因表达谱数据作为测试集数据,应用到最优预测模型中预测协同组合药物并评估模型的预测效果。我们构建的最优预测模型在训练集数据上分类错误率(Out-of-bag estimate error rate)为0.15,曲线下面积(Area Under Curve)为0.89。Connectivity Map中符合要求的药物共组成187对抗癌药物组合。我们的最优预测模型(SyDRa)预测出这187对药物组合中有28对为潜在协同组合药物。通过公共数据库进行文献检索,我们发现其中有3对药物组合已经被报道为有效的药物组合,即:azacitidine和thalidomide,imatinib和paclitaxel,streptozocin和carmustine。可见SyDRa具有较好的预测能力。综上所述,本研究分析和筛选了与协同组合药物预测有关的重要特征,其中包括药物表型数据特征和药物基因组学数据特征。同时构建了一个基于监督学习算法的协同组合药物预测模型,该模型可为大规模协同药物组合的初筛提供参考。
[Abstract]:Complex diseases such as cancer, such as cancer, are often systemic diseases caused by polygenic abnormalities. Gene mutation, copy number variation, and gene expression imbalance may be the causes or manifestations of cancer. The traditional "one drug target" drug development strategy has been proved to have some defects, such as susceptibility to drug resistance and toxic side effects. In addition, the robustness of the biological system makes it difficult for a single drug to completely inhibit the growth of tumor cells. Chemical therapy for cancer requires drugs to act on multiple targets and multiple signaling pathways, and to prevent the occurrence, transmission and action of disease related signals from multiple targets and pathways. Combined drug therapy is the most targeted therapy. A commonly used therapy has the advantages of low toxicity and high efficiency. Therefore, it is of great significance to establish a collaborative combination drug screening method. The traditional method of screening synergistic combination drugs is experimental screening. However, the experimental screening method is often costly and inefficient, and only a small amount of synergistic drug combinations can be found. The computer simulation method has the advantages of large flux, high efficiency, low cost and less pollution. With the accumulation of multi - component data and the continuous development of bioinformatics technology, the computer simulation method is expected to replace the experimental screening method for collaborative drug screening. It will be of great significance to establish an effective model for predicting anticancer synergistic drugs. In recent years, the research of collaborative drugs based on computational models has developed rapidly. According to the difference of machine learning algorithms used in building prediction models, these models can be divided into unsupervised learning prediction models and semi supervised learning prediction models. Model and supervised learning prediction model. The more often used in the supervised learning prediction model is the random forest algorithm (Random Forest). The model based on the random forest algorithm requires a proper number of labeled samples to do the training set data, then the feature design, and then the marked Yang Benjin line characteristic value score. The core of these prediction models is the These characteristics are often designed based on drug phenotypic data, such as drug targets, chemical structures, etc. Although these characteristics are also important characteristics of the drug, these characteristics do not fully reflect the pharmacological properties of drugs. Pharmacogenomics data (the expression profiles after the drug processing cell lines) can be accurately measured. Reflecting the disturbance of the drug on the cell line, it can reflect the pharmacological properties of the drug to a certain extent. Therefore, when the characteristics are designed, the drug phenotypes and pharmacogenomics features need to be taken into consideration. This study uses the known synergistic drug combinations in DREAM Challenge 7 sub-challenge 2, superimposed drug combinations and antagonistic combinations of drugs The characteristics of the 21 combined drugs (including drug phenotypes and pharmacogenomics) were designed and analyzed on the basis of the corresponding drug phenotypic data and the gene expression profiles after the treatment of the cell lines. The optimal prediction model was obtained by using the random forest algorithm for feature selection and model construction. Then we screened the model. The anti cancer drugs approved by the US Food and Drug Administration (U.S.FDA) in the Connectivity Map database and the data of the phenotypic data of these drugs and their corresponding gene expression profiles are used as test set data to predict the synergistic combination drugs in the optimal prediction model and evaluate the prediction effect of the model. The classification error rate (Out-of-bag estimate error rate) on the training set is 0.15, and the area under the curve (Area Under Curve) constitutes a total of 187 pairs of anticancer drugs in 0.89.Connectivity Map. Our optimal prediction model (SyDRa) predicts that 28 pairs of these 187 pairs of drug combinations are potential synergistic combinations. Bibliographic retrieval through public databases, we found that 3 of these combinations have been reported as effective drug combinations, that is, azacitidine and thalidomide, imatinib and paclitaxel, streptozocin and carmustine. can be found to have good predictive ability. In summary, this study analyzed and screened synergistic drugs. The important features of the prediction are the characteristics of drug phenotypic data and pharmacogenomics data, and a collaborative combined drug prediction model based on supervised learning algorithm is constructed, which can provide reference for the screening of large-scale synergistic drug combination.
【学位授予单位】：上海海洋大学
【学位级别】：硕士
【学位授予年份】：2017
【分类号】：R96

【相似文献】