This paper introduces Hierarchical Molecular Graph Self-supervised Learning (HiMol), a pre-training framework for learning molecular representations to predict properties. HiMol uses a Hierarchical Molecular Graph Neural Network (HMGNN) to encode motif structures and extract hierarchical representations. Multi-level Self-supervised Pre-training (MSP) employs generative and predictive tasks as self-supervised signals. Results on classification and regression tasks demonstrate HiMol's effectiveness, and visualizations show its ability to capture chemical semantic information.

Xuan Zang, Xianbing Zhao, Buzhou Tang