From Structure to Semantics: Training-Free LLM Guidance for Biomedical Knowledge Graph Fusion

Introduction: Entity alignment and link prediction in biomedical knowledge graphs are challenged by sparsity, heterogeneity, and evolution, requiring methods that work in cold-start settings without retraining. We introduce LLM-KG2F, a training-free framework that leverages Large Language Models (LLMs) for semantic reasoning to overcome these issues. Materials and Methods: LLM-KG2F combines BootEA-TransH structural embeddings with a zeroshot LLM plausibility prior. Without fine-tuning the LLM, it uses a chunk–reason–verify inference procedure. We evaluated the framework on three biomedical knowledge graphs: MED-BBK-9K, CKG, and RNA-KG. Results: LLM-KG2F demonstrated strong link prediction performance, achieving Hits@10 scores of 88.78 on MED-BBK-9K and 83.92 on CKG. It also produced competitive results on the RNA-KG interaction graph. Discussion: By injecting a training-free LLM prior into a structural fusion pipeline, LLM-KG2F effectively addresses the challenges of data sparsity and imbalance in biomedical knowledge graph fusion. The method's ability to operate in a cold-start setting makes it particularly suitable for the dynamic and evolving nature of biomedical data, eliminating the need for constant retraining. Conclusion: By integrating a training-free LLM prior into a structural fusion pipeline, LLM-KG2F effectively addresses data sparsity and supports cold-start link prediction. This makes the framework a robust and scalable solution for intelligent analysis in dynamic biomedical domains, eliminating the need for constant retraining.