Enhancing biomedical relation extraction with directionality.

Area of Science:

• Biomedical informatics
• Natural Language Processing
• Computational Biology

Background:

• Biological relation networks are crucial for understanding complex biological mechanisms.
• The rapid expansion of biomedical literature makes updating these networks challenging.
• Existing datasets like BioRED facilitate automated relationship extraction but lack entity role directionality.

Purpose of the Study:

• To address the lack of directionality in the BioRED corpus.
• To develop a novel multi-task language model for joint relationship, novel finding, and entity role identification.
• To enrich the BioRED corpus with essential directionality annotations.

Main Methods:

• Annotated entity roles for relationships within the BioRED corpus.
• Developed a novel multi-task language model incorporating soft-prompt learning.
• Evaluated the model's performance on relationship and entity role identification tasks.

Main Results:

• Created an enriched BioRED corpus with 10,864 directionality annotations.
• The proposed multi-task model demonstrated superior performance compared to state-of-the-art large language models (GPT-4, Llama-3).
• Successfully identified relationships, novel findings, and entity roles with high accuracy.

Conclusions:

• The enriched BioRED corpus provides crucial directionality for biological network analysis.
• The novel multi-task model effectively extracts complex biological relationships and entity roles.
• This work advances automated knowledge discovery in biomedical literature.