FGAugGCL: Molecular Contrastive Learning with Functional Group-Guided Augmentation

Molecular pre-training representations play a crucial role in modern drug discovery. Graph Contrastive Learning (GCL) has become a widely adopted pre-training approach. However, many existing GCL methods often rely on random graph augmentation, which leads to augmented views lacking logical coherence and semantic consistency, increasing the risk of the model learning incorrect knowledge. Furthermore, the critical functional group (FG) information of molecules is seldom explored in current GCL methods. To address this, a framework called Functional Group-Augmented Graph Contrastive Learning (FGAugGCL) is proposed, which integrates functional group (FG) priors into graph augmentation. Specifically, we first design the FG-guided augmentation strategy to guide augmented views in retaining essential chemical semantic information. Then, a Global-Local Attention Pooling (GLAP) method is introduced to facilitate comprehensive learning of molecular properties. GLAP combines node-level and graph-level features. It uses an adaptive attention mechanism to dynamically adjust each node’s contribution and to capture both overall structure and key local information. We evaluate FGAugGCL on various datasets. Experimental results demonstrate its superiority over state-of-the-art baselines, particularly achieving an average 6.25% reduction in RMSE on regression tasks.