Abstract
<jats:p>Predicting molecular properties from molecular graphs is a central task in cheminformatics and computer-aided drug discovery, yet the scarcity of labeled molecules limits the generalization of supervised graph neural networks. Self-supervised and contrastive pre-training methods have been introduced to mitigate this label scarcity, but they typically adopt a fixed difficulty schedule for positive and negative pairs and fail to exploit the hierarchical structural complexity of molecules, resulting in limited discriminative power on hard molecules. In this paper, we propose MolCurriCL, a molecular property prediction framework that integrates self-supervised pretraining, curricular contrastive regularization, and a local-non-local attention encoder. MolCurriCL first performs functional-group-level masked reconstruction and subgraph contrastive pretraining on a large corpus of unlabeled molecules. It then introduces a curricular contrastive regularization module that schedules positive and negative molecule pairs from easy to hard according to a structural complexity score combining the number of atoms, bonds, rings, and stereocenters, using an InfoNCE loss with a stage-wise increasing margin. To capture both local functionalgroup substructures and non-local long-range interactions, we design a local-non-local attention encoder in which local attention focuses on one-hop neighborhoods and non-local attention models cross-ring interactions through a gated fusion. We further introduce an adversarial molecular augmentation strategy that synthesizes hard negative molecules to sharpen the contrastive objective. On the MoleculeNet benchmark, MolCurriCL achieves an average AUROC of 81.7 percent on five classification datasets, outperforming the strongest baseline by 1.7 percentage points, and obtains the lowest RMSE on three regression datasets. Extensive ablations confirm the contribution of each component.</jats:p>