MolBioKG grounds unregistered molecules in biomedical knowledge graphs via structural anchors

Biomedical knowledge graphs speed drug discovery, but standard pipelines assume query molecules already exist as graph entities, leaving unregistered compounds disconnected. An arXiv preprint introduces MolBioKG, a two-layer system that grounds those out-of-graph molecules through multi-resolution structural anchoring. MolBioKG links an index of 2.74 million molecules, represented by scaffolds, fragments, functional groups, and fingerprints, to a knowledge graph with 9.6 million edges. From a SMILES string alone, it retrieves structurally related graph entities and traverses their biomedical neighborhoods without task-specific training. Inference uses static multi-anchor retrieval with Reciprocal Rank Fusion, plus Adapt-KG, a tool-using large language model policy for adaptive traversal. On the authors' benchmarks covering in-graph link recovery, multi-hop reasoning, and out-of-graph generalization, MolBioKG beat strong baselines. Hits@10 rose from 0.585 to 0.876 on multi-hop reasoning, and out-of-graph target recall rose from 0.145 to 0.269. The system keeps predictions tied to structural anchors and source-attributed graph evidence.