Dual-Branch Graph Neural Network for Predicting Molecular Odors and Discovering the Relationship Between Functional Groups and Odors

Abstract

Prediction of molecular odors is crucial for synthetic chemistry and the perfume industry. This paper presents a dual-branch graph neural network model for predicting molecular odors, named ScentGraphX, which combines transfer learning with graph attention mechanisms to address limitations of existing models. The ScentGraphX model captures atomic, chemical bond, and structural features of molecules through a feature encoder and a subgraph encoder. Experimental results show that the ScentGraphX model exhibits superior performance on a dataset comprising 4967 molecules, accurately predicting multi-label odor descriptors of molecules. Comparative analysis demonstrates that the ScentGraphX model excels in precision, recall, F1, and AUCROC evaluation metrics, validating its effectiveness in the field of molecular odor prediction. Moreover, interpretability analysis of the model reveals the impact of various chemical functional groups on odor characteristics, and ablation studies confirm the indispensability of each module in ScentGraphX.