Scaffold Data Augmentation for Molecular Property Prediction

Recently, as the applications of machine learning boom in biochemistry, data augmentation has demonstrated its power in molecular generation tasks. Specifically, data augmentation can effectively relieve the problems that insufficient training data results in model overfitting in molecular property prediction, etc. While existing works focus on the rationality of augmented construct but neglect the importance of molecular scaffolds for the task of molecular property prediction. This paper analyzes the contribution of scaffolds in property prediction tasks and proposes a new augmentation technique that preserves functional groups and modifies molecular scaffolds during the augmentation process. By modifying scaffolds, data augmentation can increase the diversity of molecules and thus enrich the dataset. At the same time, by preserving the functional groups, the introduction of noise can be effectively reduced, the quality of augmented data can be improved, and the invariance of labels can be enhanced. We conducted experiments on four benchmark datasets using three baselines with different classification models to test the effectiveness of our proposed method. Our results strongly demonstrate that data augmentation with modifying scaffolds can effectively optimize property prediction performance and improve model generalization.