Curriculum-aware Training for Discriminating Molecular Property Prediction Models

Molecular property prediction plays a crucial role in various fields such as cheminformatics and artificial intelligence. Despite its wide applicability, current models still struggle in the presence of activity cliff, in which molecules with similar chemical structures display remarkable different properties. This hinders the model's ability to learn distinctive representations for molecules with similar chemical structures, resulting in inaccurate predictions on molecules with activity cliff. In this paper, we first present empirical evidence demonstrating the ineffectiveness of standard training pipelines on these molecules. We then propose a novel approach that reformulates molecular property prediction as a node classification problem, and introduce both node-level and edge-level tasks to improve the learning for these challenging molecules. The proposed method is versatile, and can be seamlessly integrated into a variety of pre-trained or randomly initialized base models. Extensive evaluation on various molecular property prediction datasets validate the effectiveness of our approach. Published as a conference paper at ICLR 2025 minor differences (the two yellow boxes), but their responses to the ER, ATAD5 and HSE receptors are all different.