基于随机森林和深度神经网络的药物ADMET性质预测  

Prediction of Drug ADMET Properties Based on Random Forest and Deep Neural Network

作　　者：王肖成 阮昊 鹏奕锟[1] 李成堃 陈雪 WANG Xiaocheng;RUAN Hao;PENG Yikun;LI Chengkun;CHEN Xue(College of Big Data and Information Engineering,Guizhou University,Guiyang 550025,China;Shanghai Institute of Optics and Fine Mechanics,Chinese Academy of Sciences,Shanghai 201800,China;Shanghai Juji Technology Co.,Ltd.,Shanghai 200072,China;School of Medicine,Guizhou University,Guiyang 550025,China)

机构地区：[1]贵州大学大数据与信息工程学院,贵阳550025 [2]中国科学院上海光学精密机械研究所,上海201800 [3]上海聚迹科技有限公司,上海200072 [4]贵州大学医学院,贵阳550025

出　　处：《微处理机》2023年第2期39-43,共5页Microprocessors

基　　金：国家自然科学基金(61264004);贵州省高层次创新型人才培养项目(【2015】4015)。

摘　　要：针对抗乳腺癌药物研发中ADMET占据比例过大且难以准确预测等问题,提出一种基于随机森林及深度神经网络模型的候选药物ADMET性质预测模型。模型在保留分子结构信息的前提下,能够减少特征冗余和样本维度,以随机森林算法及特征选择过程结合变量重要性评分方法,获取最优分子描述符变量特征。通过改进DNN模型结构中各层之间的快捷连接方式,更有效地保留分子结构信息。在公开数据集中进行对比试验,结果表明所模型在测试集中5种ADMET性质分类预测平均准确度可达89.15%,优于当前主流模型,具有更强的适用性。Aiming at the problem that ADMET occupies a large proportion in the research and deve-lopment of anti-breast cancer drugs and is difficult to predict accurately,a prediction model of ADMET properties of candidate drugs based on random forest and deep neural network model is proposed.The model can reduce feature redundancy and sample dimension while retaining molecular structure infor-mation.Random forest algorithm and feature selection process combined with variable importance scoring method are used to obtain the optimal molecular descriptor variable features.By improving the quick connection mode between layers in DNN model structure,the molecular structure information can be retained more effectively.A comparative experiment was carried out in an open data set,and the results showed that the average accuracy of the classification prediction of five ADMET properties in the test set could reach 89.15%,which is better than the current mainstream model and have stronger applicability.

关 键 词：随机森林 深度神经网络 分子描述符变量 ADMET性质 药物研发 

分 类 号：TP391[自动化与计算机技术—计算机应用技术] R961[自动化与计算机技术—计算机科学与技术]