A Property-Aware Framework for Molecular Property Prediction Via Substructure Contrastive Learning

Drug development processes face high failure rates and cost challenges, necessitating efficient molecular property prediction methods. This research proposes a property-aware substructure feature fusion algorithm to address two key limitations in existing models: the neglect of cross-modal semantic alignment at the substructure level and the lack of mechanisms for dynamically adjusting substructure importance across different properties. To overcome these limitations, we innovatively extend cross-modal contrastive learning to the functional group level, achieving effective semantic alignment between molecular functional groups and their textual descriptions. Furthermore, we introduce a property-aware cross-attention mechanism capable of dynamically evaluating contribution weights of various substructures based on specific properties. Experimental validation across four public datasets demonstrates significant performance improvements, confirming the effectiveness and applicability of our method in molecular property prediction tasks relevant to drug discovery.