Abstract
The accurate prediction of pure component properties has consistently been a critical issue in fields such as chemical engineering, biomedicine, and environmental science. In recent years, end-to-end deep learning methods have shown significant improvement over traditional machine learning approaches. This is due to their ability to automatically learn task-relevant representations from raw molecular data. In addition to accurate property prediction, researchers have increasingly focused on how specific fragment structures influence molecular properties. However, existing fragmentation methods based on predefined rules and group libraries struggle to capture novel molecular structures, which hampers the development of new materials and drugs. To address these challenges, this work proposes a hierarchical molecular fragmentation method. This method can automatically segment molecules into multiple fragments containing key functional groups. Then a three-branch graph attention network was constructed to achieve multi-level representation. Finally, a multi-layer perceptron is employed to establish the mapping relationship between molecular features and physical property values. Twenty datasets were used for validation, which can be grouped into four categories: Thermodynamic Properties, Pharmacokinetics, Toxicological Properties, and Industrial Safety. The results show that the best performance is achieved, with the average error reduced by 6.8% compared to existing research.