Research On Odor Perception Based On Molecular Physicochemical Characteristics

With the development of digital technology,the research on other human senses such as vision and hearing is becoming more and more mature,and people have already realized "Clairvoyance" and "Omniscient".However,the development of smell is still in the exploratory stage,and there is still a long way to go before "Fragrant Miles".Olfactory perception is a way to describe the sense of smell,which uses odor descriptors to describe the molecular odor to represent the sense of smell.Due to the complexity of odor composition,people have not been able to establish a good mapping relationship between the properties of odor molecules and olfactory perception.This thesis attempts to start from the perspective of the physicochemical characteristics of odor molecules,not only including the extraction of molecular descriptor information,but also for the first time by constructing the two-dimensional plane structure and three-dimensional spatial structure of molecules into a graph structure,thereby extracting molecular graph structure information as molecular physicochemical features.Then,machine learning and deep learning are used to build a good odor perception evaluation model,so as to achieve accurate prediction of odor perception.Firstly,starting from the introduction of the collection and preprocessing of the experimental dataset,this thesis introduces the source,collection and preprocessing methods of the odor perception scoring dataset in detail.On this basis,we focus on how to construct the odor perception scoring space,the meanings of odor molecules in different representations,and their mutual conversion process.Secondly,the molecular descriptor information based on molecular physicochemical features,including molecular qualitative descriptors and molecular fingerprints,is investigated.And using a variety of machine learning methods such as random forest,ridge regression,partial least squares regression and deep neural network methods,a mapping model from molecular descriptors to molecular odor perception is constructed.In addition,by comparing the prediction effect of different molecular descriptor information and the performance of the odor perception prediction model,the feasibility and accuracy of the experiment are verified.After that,in order to further improve the accuracy of molecular odor perception prediction,this thesis proposes to use molecular graph structure information,especially three-dimensional spatial structure information,as molecular physicochemical features.Two odor perception prediction models,Odor-2DGCN and Odor-3DGCN,are built based on graph convolutional networks.In the process of extracting molecular graph structure information,this thesis also screened out 19 atomic features that are most beneficial to molecular odor perception prediction through feature selection based on correlation.The final test results show that the prediction of molecular odor perception based on molecular graph structure information is feasible and the results are good.It predicts an odor perception score of 0-100,with a minimum MAE of 1.276.Finally,this thesis compares and analyzes the results of odor perception prediction by comparing different molecular physicochemical features,namely molecular descriptors and molecular graph structure,and concludes that the proposed method uses molecular graph structure information as molecular physicochemical features to conduct experiments,which can improve the accuracy of odor perception prediction.Among them,the Odor-3DGCNmax model based on molecular three-dimensional spatial graph structure information and integrating atomic information with graph convolution maximum pooling mechanism can obtain the best odor perception prediction effect.It proves the superiority and feasibility of the method proposed in this thesis,and gives the use value and practical significance for the follow-up research and application.