The human-AI collaborative framework is now widely used in drug discovery, leveraging the strengths of both human expertise and AI to significantly speed up the development of new drugs. GNN-based methods have been applied to molecular representation learning, but few fully exploit the hierarchical structural information of molecules. Molecules consist of structural backbones and key fragments, such as functional groups ‘-OH’ and ‘-COOH’, which are crucial for determining properties like hydrophilicity. In this work, we introduce the Dual-Level Molecule Representation Learning (DLMRL) framework. Initially, the BRICS (Breaking of Retrosynthetically Interesting Chemical Substructures) method is employed to uniquely decompose the molecular graph into substructures according to 16 chemical bond-breaking rules. DLMRL performs molecular representation learning at both the entire graph and substructure levels. Using the BRICS masking method, molecular fragments serve as masking units for representation augmentation, allowing the model to focus on both the entire molecule and smaller, informative fragments. At the BRICS level, a contrastive learning strategy enhances the model’s ability to differentiate important substructures. Experiments validate DLMRL’s effectiveness and interpretability. Our source code is publicly availiable at https://github.com/gyanlala/causal Brics.