MolCHG: Multi-Level Self-Supervised Pretraining on Compositional Hierarchical Graphs

A team of researchers has introduced MolCHG, a multi-tiered self-supervised pretraining framework aimed at predicting molecular properties. This framework employs a Compositional Hierarchical Graph, which categorizes molecular structures into four types of nodes across three semantic tiers, featuring a bond graph that runs parallel to the atom graph. This approach enhances bond data into independently evolving node representations, allowing fragment nodes to equally integrate atom-level and bond-level semantics. Additionally, three pretraining objectives tailored to specific levels are proposed, including a cross-view contrastive task for atom and bond interactions. This research tackles the shortcomings of current methods that function at a single structural level and consider bond information merely as auxiliary edge attributes.