LLM-driven Biomedical Knowledge
UniD3: Unified Drug-Disease Dataset Construction via KG-RAG
UniD³ orchestrates Llama3.3-70B with Knowledge Graph Retrieval-Augmented Generation to transform over 150,000 PubMed articles into structured, high-fidelity biomedical datasets. Our dual-stage entity extraction pipeline ensures consistent, noise-resistant graph construction.
150k+
PubMed publications
0.80+
F1 across tasks
0.9005
Expert F1 (DDM)
Drug-Disease Matching (DDM)
Identify high-confidence drug and disease relationships with contextual explanations grounded in PubMed evidence.
Drug Effectiveness Assessment (DEA)
Evaluate drug outcomes, effectiveness signals, and clinical directions across large-scale biomedical corpora.
Drug-Target Analysis (DTA)
Trace molecular targets, pathways, and intervention strategies by mining structured triplets.
Dual-Stage KG Construction
UniD³ first performs paper-level extraction to capture localized research context, then promotes consistent entities into a KG-level summary. LightRAG and custom prompts distill reliable triplets that fuel downstream QA and dataset generation pipelines.
Access Knowledge Graph on Zenodo