Large Language Models for Structured Extraction of Molecular Interactions from Scientific Literature

Scientific literature contains extensive knowledge about interactions between biological entities, but much of this information is expressed in unstructured text. We present an LLM-based pipeline that extracts molecular interactions from scientific publications and represents them in Biological Expression Language (BEL), with entity grounding and structural validation. In a human evaluation of 50 extracted BEL statements, 62.8% of individual ratings judged statements correct or correct but potentially more precise, and 40% of statements received positive ratings from at least four of five reviewers. We also compare the approach with INDRA on 20 sentences. Among 72 semantically distinct outputs, 53 interactions were unique to the LLM-based pipeline, 10 were unique to INDRA, seven were exact matches, and two were close matches, indicating that the approaches capture largely complementary relationships. We further demonstrate how the resulting knowledge graphs can support graph-based retrieval for downstream LLM reasoning.