Research On Odor Perception And Its Predicting Model Based On Chemical Molecular Descriptors

Odor perception refers to the process from the molecules of odorants entering the nasal cavity to the formation of olfactory perception,such as "sweet","musky",and"fishy",in the human brain.Identifying diverse odor perceptions from the physicochemical features of odorants,namely the stimulus-percept problem,is always the hot issue in the odor research community.There are three critical issues in stimulus-percept problem:the construction of physiochemical feature space of the odorants,namely,the effective representation of the odorants through the physiochemical feature information;research on the odor perception space,including the construction and structure of the odor perception space and the relationships between different odor perceptions;the construction of mapping model,that is,how to map the physicochemical features of odorants to the odor perceptions.This work mainly focuses on the odor perception and mapping model.and our research is carried out based on two public datasets:Dravnieks and DREAM.The main contents of this paper are as follows:(1)The relationships between different odor perceptions are studied.The relationships between different odor perceptions are complicated.For instance,"sweet"and "rancid" are opposite,while "sweet" and "candy taste" are much similar.Currently,less research work focuses on the relationships between different odor perceptions.Therefore,the task of the prediction of missing odor perception from the known odor perception is proposed.To this end.Random Forest(RF)is employed as the feature selection algorithm to obtain the importance index of the known odor perceptions for the prediction of the missing odor perception.According to the descending order of importance index.the known odor perceptions are sent to the Support Vector Regression(SVR)model as the inputs.Specifically,the prediction of "pleasantness" is studied thoroughly.The experimental results show that some missing odor perception could be predicted precisely while others not.It indicates that these odor perceptions are not distributed uniformly in the odor perception space.Some of them are related closely,while the others are isolated points in the odor perception space.The "pleasantness" prediction results certify that the known odor perceptions have different effects on "pleasantness"prediction.They are classified into:"positive","negative",and "neutral".The relationships between the prediction performance and the intra-class combination and inter-class combination of the known odor perceptions are investigated.The results indicate that the intra-class combinations of known odor perceptions have little effect on"pleasantness" prediction performance.For inter-class combinations of the known odor perceptions,the combination of "negative" and "positive" inputs could improve the prediction performance significantly,while the inter-class combinations with the "neutral"known odor perceptions have not much improvement for "pleasantness" prediction.(2)The reduction of Odor Vocabulary(OV).There are thousands of odor perception descriptors in OV.Unlike the enlargement of OV in recent work,we aim to reduce the OV size.To this end,the definition of Primary Odor Perception Descriptors(POPDs)is proposed,which is a subset of the OV and could represent all the odor perceptions in the OV.Concretely,a selection mechanism based on clustering algorithms with an updating strategy and multi-output regression is contrived to identify as few and stable POPDs as possible.The experimental results indicate that dozens of odor perception descriptors are redundant,identified as Non-Primary Odor Perception Descriptors(NPOPDs).Also,a mapping model from the POPDs to the Non-POPDs(NPOPDs)is established,through which the perceptual ratings of the NPOPDs could be accurately inferred from those of the POPDs.Considering that the odor perception dataset is sparse,the relationships between the sparsities and correlations of odor perception descriptor vectors and the prediction performance are studied.Experimental results show that the greater the sparsity.the worse the prediction performance.The greater the correlation is.the better the prediction performance(3)The construction of the mapping model from the mono-molecular odorants’Chemical Molecule Descriptors to the odor perceptions.Taking the CMDs features as inputs and the perceptual ratings of odor perception descriptors as outputs,two prediction patterns.Independent Prediction(IP)Union Prediction(UP),are adopted.Considering that the CMDs features are highly correlated,a Huge Kernel and Huge Stride(HKHS)convolutional neural network(CNN)is proposed for the odor perception prediction.Specifically,IP-HKHS-CNN and UP-HKHS-CNN are employed to predict the perceptual ratings of 21 odor perception descriptors of the DREAM dataset.The experimental results certify that the prediction performance of the IP-HKHS-CNN model is superior to that of the UP-HKHS-CNN model.Besides,the influence of the number of possible values of the CMDs features on the prediction performance is investigated.For the IP-HKHS-CNN,UP-HKHS-CNN,and RF models,the results imply that those CMDs features with the number of possible values less than 30 have little effect on the odor-perception prediction performance.Namely,the predicting performance remains stable.This research could be applied for the odor assessment of the products in the food,perfume industries.Besides,the mapping model from the odorants’ CMDs features to the odor perceptions might have a particular reference value for the perfume industry’s product design.