Pre-training Molecular Graph Representations with Motif-Enhanced Message Passing

Recently, molecular representation learning (MRL) based on graph neural networks (GNNs) has flourished. Many existing approaches are purely data-driven, focusing on exploiting the intrinsic topology and construction rules of molecules without any chemical prior knowledge. The high dependency on data makes it difficult for them to generalize to a broader chemical space. To address this issue, we introduce the property-aware molecular GNN approach, which, grounded in domain knowledge of chemistry, fragments molecules into motifs and then quantizes the motifs feature vectors through our proposed Discrete Motif Message Passing (DMMP) framework. This extends the representational space of motifs and enhances their generalizability. Moreover, to make the entire molecular graph aware of good motifs features, we construct an augmented molecular graph and design three levels of pretraining tasks. Extensive experiments demonstrate that our proposed multilevel self-supervised pre-training of augmented molecular graph representation learning (MSPmol) outperforms state-of-the-art baselines in molecular property prediction and drug-target affinity (DTA) tasks while providing reasonable chemical interpretations. This work aids a wide array of downstream tasks by offering high-quality molecular representations.