pipette
ENEnglish

SPIDER: Contextualizing Biomedical Knowledge for Evidence-Grounded Scientific Reasoning

M. Liu, W. Zhao, J. Chen, X. Quan, X. Wang, J. Feng, X. Zhou, J. Han, Y. Zhang, Y. Wang, X. Zhang, W. Xie

Preprint

En palabras de los autores

Aggregated relations in biomedical knowledge graphs can obscure evidence and study conditions, complicating the distinction between scientific disagreement and contextual differences. Here we introduce SPIDER, a multi-agent framework grounding biomedical relations in literature evidence and extracting study conditions for interpretation. Applied to a graph integrating 48 databases and ontologies, SPIDER analyzed 1.95 million PubMed abstracts, adding 2.31 million previously unrepresented, literature-supported entity pairs to construct SPIDER-KG. Daily updates incorporate relations and evidence from newly indexed PubMed records. Across 89,937 evidence-supported comparisons of potentially incompatible relations, SPIDER classified 79.0% as contextual differences, and 6.1% as genuine disagreement, achieving 80.3% macro-F1 averaged across both categories against expert annotations. On SPIDER-Bench and four biomedical reasoning benchmarks, SPIDER-KG improved large language models' interpretation of evidence-supported biomedical relationships. Retrospective analysis of newly reported gene--disease relationships further demonstrated how evidence-linked knowledge can contextualize emerging discoveries. SPIDER-KG provides a reusable, incrementally extensible evidence layer for biomedical reasoning.

Resultado principalEl resumen no menciona limitaciones.

Apareció: jueves, 19 de noviembre. bioRxiv. Preprint, todavía sin revisión por pares.

DOI: 10.64898/2026.11.11.765839