MetaphorPrompt2-A Structure and Function-Focused Approach for Extracting Causal Events from Biological Text

Extracting molecular regulatory pathways (MRPs) from biomedical literature is crucial for building knowledge graphs that represent disease mechanisms and molecular regulation. However, even advanced large language models (LLMs) using in-context learning often misinterpret complex domain-specific causal statements or omit intermediary steps, resulting in incomplete pathway representations. MetaphorPrompt2 is motivated by cognitive theories of structural mapping and causal event representation. It improves MRP extraction by emphasizing structural relations and functional roles of biological entities rather than relying on surface-level grammar. The system integrates 5 components that collectively reduce parsing complexity and mitigate error propagation. At one-shot in-context learning, the proposed method achieves a 31% improvement in edge prediction F1 over the no-metaphor baseline and a 6.5% to 12.2% improvement over a previous method across 3 datasets: reguloGPT, BioInfer, and ADE. The percentage of causal links where both nodes are missed is reduced from 7.7% in the previous method to 3.7% in MetaphorPrompt2. These results support improved causal triplet extraction for biomedical pathway construction and potentially downstream utility in hypothesis generation or drug repurposing, which remains to be evaluated.