基于PaDEL分子描述符的气味机器学习识别研究  

Odor molecule recognition via PaDEL molecular descriptors and machine learning

作　　者：苏洋洋 夏仁杰 王云松 许振军 SU Yangyang;XIA Renjie;WANG Yunsong;XU Zhenjun(School of Electrical and Information Engineering,Jiangsu University of Technology,Changzhou 213001,China;Zhejiang Guyue Longsan Electronic Technology Development Co.LTD,Shaoxing 312000,China)

机构地区：[1]江苏理工学院电气信息工程学院,江苏常州213001 [2]浙江古越龙山电子科技发展有限公司,浙江绍兴312000

出　　处：《江苏理工学院学报》2024年第2期5-10,共6页Journal of Jiangsu University of Technology

基　　金：常州市基础研究计划(应用基础研究)项目“人工智能辅助筛选新冠病毒S蛋白与宿主ACE2蛋白结合抑制剂”(CJ20200045);江苏省青年自然科学基金项目“SIRT1激活的分子机理研究及新型别构激动剂筛选”(BK20191032)。

摘　　要：气味识别有助于更好地理解嗅觉处理机制和生物进化中的交流方式,也是食品、饮料和香水等商业领域的关键技术。在利用PaDEL对气味分子进行数值化表征的基础上,采用决策树、随机森林、极端随机树、支持向量机、全连接神经网络和循环神经网络等机器学习算法,对PaDEL分子描述符进行了特征学习,以实现对气味的多标签分类。结果显示,FCNN和DT模型的预测性能较强,BiGRU模型的预测性能最弱。采用主成分分析算法对PaDEL数据进行了降维处理,基于此数据的不同机器学习模型气味分类性能比较中,BiGRU模型的性能最优。Odor recognition contributes to a better understanding of olfactory processing mechanisms and com-munication modes in biological evolution,and is a key technology in various commercial sectors such as food,beverages,and perfumes.Based on the numerical characterization of odor molecules using PaDEL,feature learn-ing of PaDEL molecular descriptors is carried out by using machine learning algorithms such as decision tree,random forest,extreme random tree,support vector machine,fully connected neural network and recurrent neu-ral network to realize multi-label classification of odor.The results show that the predictive performance of FCNN and DT models is relatively good,while the predictive performance of BiGRU model is worst.The principal com-ponent analysis was utilized to reduce the dimensionality of the PaDEL data,and in the comparison of odor classi-fication performances of different machine learning models based on this data,BiGRU model is the best.

关 键 词：PaDEL分子描述符 机器学习 气味识别 

分 类 号：TP183[自动化与计算机技术—控制理论与控制工程]