Study Of Molecular Taste/Odor Prediction Model And Quantitative Structure-Activity Relationship Based On Deep Learning

In our daily life,the perception of "taste" was produced by senses of taste and odor.Taste molecules and odorants were not only used to adjust and improve the tastes of food,but also closely connected with human health.In terms of taste and odor identification,although experimental methods can get more accurate results,the process was time-consuming and laborious,and cannot be identified in large quantities;traditional machine learning methods were widely used in taste and odor prediction,but the accuracy of prediction results was relatively poor;deep learning methods have been used for taste and odor prediction by virtue of their ability to build high-performance models.At present,the problems of deep learning in molecular taste and odor prediction were as following:(1)the use of manual screening of features leads to large errors;(2)it is difficult to explain the molecular quantitative structure-taste/odor relationship.To this end,this research carried out the following work:1)This paper used Multilayer Perceptron(MLP)and molecular fingerprints,MLP and molecular descriptors,as well as Convolutional Neural Network(CNN)and molecular two-dimensional images to construct to 9 models for molecular bitterant/nonbitterant,sweetner and non-sweetner,as well as bitterant/sweetner predictions,and then studied the Quantitative Structure-Taste Relationship(QSTR).The results showed that these 9 models have their own advantages in molecular taste prediction.Among them,these models based on MLP and molecular fingerprints had the strongest predictive ability(AUC values: 0.94,0.94 and 0.95,respectively).The model built based on CNN and molecular images only took simple two-dimensional molecular images as input.CNN models realized automatic feature extraction to avoid errors caused by manual feature selection and had strong predictive capability(AUC values: 0.88,0.90,and 0.91,respectively).Models constructed by MLP and molecular descriptors also had strong predictive power(AUC values: 0.94,0.84,and 0.87,respectively).The models based on MLP and molecular descriptors explained the QSTR of molecules.The structural characteristics that were conducive to recognition bitterants included molecular hydrophobic characteristics,molecular composition information and molecular shape.Characteristics that were benificial to recognition of sweetners included molecules surface interaction,molecular weight and electrical charge.2)Meanwhile,based on MLP and molecular descriptors,MLP and molecular fingerprints,as well as CNN and molecular two-dimensional images,this paper constructed 12 binary-class prediction models for odor/odorless,fruity/odorless,floral/odorless and woody/odorless,and 3 multi-class prediction models for odorless/fruity/floral/woody.Then the molecular Quantitative Structure-Odor Relationship(QSOR)was studied.The models that constructed by MLP and molecular descriptors exhibited the highest predictive ability(The AUC values of all two-class prediction models were 0.99 and the AUC value of the multi-class prediction model was0.86).The CNN models required only simple input,which realized the automatic extraction of the molecular features,and had a strong predictive ability(The AUC values of all binary-class prediction models were over 0.98 and the AUC value of the multi-class prediction model was 0.80).The models based on MLP and molecular fingerprints also had favorable results(The AUC values of all binary-class prediction models were greater than 0.93 and the AUC value of the multi-class prediction model was 0.80).The models based on MLP and molecular descriptors were used to study the QSOR of molecules.The results showed that molecular weight,electronegativity,and surface interaction were conducive to identifying odorants;partial charge and electronic information,molecular surface interaction and molecular shape were conducive to identifying fruity molecules;molecular hydrophobic characteristics,molecular composition information and electronic information were benificial to identifying floral molecules.The molecular surface interaction,electronic information and molecular shapes were conducive to the recognition of woody molecules.In summary,the models based on MLP and molecular fingerprints and the models based on MLP and molecular descriptors had the strongest molecular taste and odor prediction ability,respectively.In this paper,the CNN model was used for the first time to predict molecular bitterant and sweetner.The CNN models realized molecular taste and odor prediction by virtue of molecular two-dimensional images,avoiding the errors caused by manual feature selection,and obtaining better results.In addition,the models that built on MLP and molecular descriptors obtained features that were beneficial to identifying bitterant,sweetner,and multiple odor molecules.These models constructed in this research were helpful for molecular taste and odor prediction.The QSTR and QSOR investigation help understand the generic mechanism of the relationship between chemical structure and taste/odor perception,which provides a theoretical basis for the interpretability research of deep learning.