MolBioKG Grounds Unseen Molecules in Biomedical Knowledge Graphs

Biomedical knowledge graphs (KGs) are invaluable for drug discovery, but their utility is limited when encountering molecules not already registered within the graph. This "out-of-graph molecule problem" leaves novel or unseen compounds disconnected from existing biomedical evidence. To overcome this, researchers introduce MolBioKG, a two-layer system designed to ground these unregistered molecules. MolBioKG achieves this by employing multi-resolution structural anchoring, connecting an index of 2.74 million molecules (represented by scaffolds, fragments, functional groups, and fingerprints) to a vast 9.6-million-edge KG. Given only a SMILES string, the system can retrieve structurally related graph entities and traverse their biomedical neighborhoods without requiring task-specific training. It features both static multi-anchor retrieval and an LLM-based adaptive traversal policy, significantly outperforming baselines in link recovery, multi-hop reasoning, and out-of-graph generalization, while maintaining traceable structural anchors and source-attributed KG evidence.