Research On Odor Substance Properties Based On QSAR Models

In recent years,the prediction of odorant substance properties,as an important and popular research topic in the olfactory system,has received more and more attention in academia and industry.However,it is difficult to predict the key properties of odorant substance from the formulation during the development of a flavor or fragrances,and relying on traditional methods for determination has certain limitations,such as long time periods and high costs.Therefore,how to construct a more efficient and low-cost way for predicting the properties of odorant substance presents a new challenge.Quantitative structure-activity relationship（QSAR）theory,as an important tool to study the structure and activity of compounds,plays a powerful role in the odorant substance properties,and provides a new direction for rapid and precise blending into new flavors and fragrances.In this thesis,the relevant monomer flavors datasets were collected and organized,and the effective odor representation information was calculated by using structural informatics software,and constructed theoretical prediction models for fragrance retention grades and odor intensity activity,and discussed the performance and applicability of different model algorithms.The main research contents of this thesis are as follows:（1）Research on the prediction method of fragrance retention grades of monomer flavors based on QSAR model.Firstly,1552 monomer flavors datasets containing fragrance retention grades（FRGs）activities were collected and sorted,and the structural information was calculated by Dragon 7.0 software.Then,five variable selection techniques including Principal Component Analysis,Lasso,Recursive Feature Elimination,Autoencoder and Boruta algorithm were employed to reduce the model training cost.Next,a series of QSAR prediction models were constructed by Random Forest,Support Vector Machine and Deep Neural Network algorithm.On this basis,we developed a weighted scoring formula to calculate the correlation score between functional groups and FRGs.The results demonstrated that SH（thiols）,ArOR（ethers）,and ArCOOR（esters）functional groups have significant effect on the FRGs.Furthermore,we defined the applicability domain to limit the application scope of the test dataset and used external data to verify the model reliability.After comparing analysis using the Recursive Feature Elimination to obtain 80-dimensional effective molecular descriptors,the Random Forest algorithm has the best performance,with an accuracy of77.81%,a precision of 77.83%,a recall rate of 77.99%,and an F1-score of 77.88%.The classification model can be employed as a reliable tool for the evaluation of the FRGs attributes of monomer flavors and provides the theoretical basis and technical support for the precise design,intelligent screening,and efficient synthesis of new flavors or fragrances.（2）Research on the prediction method the odor intensity of monomer flavors based on POI-3DGCN.Firstly,we proposed a novel and efficient framework（POI-3DGCN）that combines a Graph Convolutional Neural Network algorithm and the 3D topology of odor molecules based on 1200 datasets of monomer flavors containing odor intensity activity.Compared with five baseline models,Random Forest,Support Vector Regression,Multi-Layer Perceptron,Long-Short-Term Memory network and Graph Attention Network,the performance of the 3DGCN model on the task of odor intensity activity prediction is significantly improved.Secondly,we analyzed the impact of different pooling mechanisms on the performance of the 3DGCN model.The results demonstrateed that global aggregation set2 set pooling operation to improve the performance of 3DGCN model predicting odor intensity on monomer flavors datasets with sparse graph structure.More significantly,we found that the rotations of odor molecules in 3D space have the same topology,except for the3 D orientation,and the 3DGCN model predicts odor intensity with has rotation equivariance,which also reflects the good robustness of the model.Finally,we visualized the generation process of molecular features as well as the contribution of atomic features distributed over molecules to explore the interpretability of the model.The POI-3DGCN framework proposed in this thesis is reliable in evaluating the odor intensity activity of monomer flavors,and lays a solid theoretical foundation for the realization of 3D features in the field of deep learning olfactory perception.Finally,the work done in this thesis is briefly summarized,and the future research work is prospected on the basis of summarizing the full text.