PACL: Property-Aware Contrastive Learning with Adaptive Substructures for Molecular Property Prediction

Molecular property prediction is a central task in drug discovery, yet acquiring labeled data remains costly and time-consuming. Self-supervised contrastive learning provides a promising route for learning from abundant unlabeled molecular data. However, current contrastive methods still face two challenges: constructing chemically meaningful positive and negative pairs, and adapting shared molecular representations to task-specific property signals. To address these limitations, we propose a Property-Aware Contrastive Learning (PACL) framework for molecular property prediction with adaptive substructures. PACL introduces a cross-scale contrastive learning strategy that aligns atomic-level representations with adaptively partitioned substructure-level embeddings, avoiding reliance on data augmentation or 3D conformer generation. During fine-tuning, task-specific learnable prototypes and a property-aware embedding module recalibrate molecular representations to emphasize property-relevant features. Evaluated on nine MoleculeNet benchmarks under scaffold splitting, PACL achieves the best average ROC-AUC for classification and the lowest average RMSE for regression among the compared methods. Visualization and prototype preference analyses further indicate that PACL separates structurally similar molecules with different functional properties more clearly in task-aligned representation spaces. These results support adaptive substructure contrastive learning and task-specific semantic alignment as an effective strategy for multi-task molecular property prediction in data-limited settings.