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

Area of Science:

• Bioinformatics
• Computational Biology
• Natural Language Processing

Background:

• Extracting molecular regulatory pathways (MRPs) is vital for understanding disease mechanisms and building biomedical knowledge graphs.
• Current large language models (LLMs) struggle with complex causal statements in biomedical literature, leading to incomplete pathway representations.

Purpose of the Study:

• To develop an improved method for extracting molecular regulatory pathways (MRPs) from biomedical text.
• To address the limitations of existing LLMs in interpreting complex causal relationships and representing biological pathways accurately.

Main Methods:

• Introduced MetaphorPrompt2, a system inspired by cognitive theories of structural mapping and causal event representation.
• Emphasized structural relations and functional roles of biological entities over surface-level grammar.
• Integrated 5 components to reduce parsing complexity and mitigate error propagation.

Main Results:

• Achieved a 31% improvement in edge prediction F1 score over a no-metaphor baseline in one-shot learning.
• Demonstrated a 6.5% to 12.2% improvement over a previous method across three datasets (reguloGPT, BioInfer, ADE).
• Reduced the omission of causal links (both nodes missed) from 7.7% to 3.7%.

Conclusions:

• MetaphorPrompt2 significantly enhances causal triplet extraction for constructing biomedical pathways.
• The method shows potential for downstream applications such as hypothesis generation and drug repurposing.
• Future work should evaluate the practical utility of these improved pathway representations.