MolBioKG: Grounding Out-of-Graph Molecules in Biomedical Knowledge Graphs
MolBioKG is a cutting-edge system aimed at solving the cold-start problem faced by unregistered molecules in biomedical knowledge graphs (KGs). Unlike typical approaches that assume the queried molecules are already in the graph, MolBioKG successfully links these unseen molecules. It achieves this through multi-resolution structural anchoring, connecting to a KG that includes 9.6 million edges and 2.74 million indexed molecules, depicted as scaffolds, fragments, functional groups, and fingerprints. The system can retrieve related graph entities from a SMILES string and explore their biomedical contexts without needing specific training. It uses two inference methods: static multi-anchor retrieval with Reciprocal Rank Fusion and Adapt-KG, which is a tool-using LLM policy for dynamic traversal. Evaluations show improved performance on various tasks. The research is available on arXiv under identifier 2608.06713.
Key facts
- MolBioKG is a two-layer system for grounding unseen molecules in biomedical knowledge graphs.
- It connects an index of 2.74 million molecules to a 9.6-million-edge KG.
- Molecules are represented by scaffolds, fragments, functional groups, and fingerprints.
- Input is a SMILES string; no task-specific training is required.
- Two inference mechanisms: static multi-anchor retrieval (Reciprocal Rank Fusion) and Adapt-KG (tool-using LLM policy).
- Evaluated on in-graph link recovery, complex multi-hop reasoning, and out-of-graph tasks.
- Paper available on arXiv:2608.06713.
Entities
Institutions
- arXiv