POI-3DGCN: Predicting odor intensity of monomer flavors based on three-dimensionally embedded graph convolutional network

The relationship between spatial information of molecular topology and odor characteristics has received increasing attention from the academic community. Odor intensity (OI) prediction, as one of the most important topics in the olfactory systems (OS), has been thoroughly investigated in the past. Nonetheless, traditional methods have certain limitations as they usually require high-precision instruments and are time-consuming in collecting OI datasets. Here, we created a novel and efficient framework (POI-3DGCN) for OI prediction based on a three-dimensional graph convolutional network (3DGCN) model to overcome these challenges. Compared with other advanced models (RF, SVM, MLP, LSTM and GAT), the 3DGCN model exhibits significantly higher performance on the task of predicting OI. We confirmed that global pooling (aggregation), especially s e t 2 s e t , improves the predicted OI performance on monomer flavor datasets with sparser graph structures. More significantly, we found that the rotations of odor molecules in 3D space have the same topology, except for their 3D orientation, and the 3DGCN model predicts OI with has rotation equivariance, which also reflects the good robustness of this model. Consequently, the proposed POI-3DGCN model is likely to be considered reliable for evaluating the OI of monomer flavors, which lays a good foundation in the implementation of three-dimensionality to the field of deep learning olfaction. • 3DGCN-based default prediction model is proposed. • The contribution of atomic features to predicting the OI of monomer flavors in 3DGCN. • Global pooling based on iterative content-based attention improve model performance. • Application of monomer flavor data shows that our model outperforms existing model.