BioREx: Improving Biomedical Relation Extraction by Leveraging Heterogeneous Datasets.

Area of Science:

• Biomedical Natural Language Processing (NLP)
• Information Extraction
• Computational Biology

Background:

• Biomedical relation extraction (RE) is crucial for literature-based discovery and knowledge graph construction.
• Current RE models are limited by small, domain-specific datasets and expensive manual annotation.
• Developing generalized and high-performing RE models is challenging due to data heterogeneity.

Approach:

• A novel framework was developed to systematically address data heterogeneity and combine individual RE datasets into a large, unified dataset.
• BioREx, a data-centric approach, was implemented using the combined dataset for relation extraction.
• The framework and dataset were packaged as a stand-alone tool for wider accessibility.

Key Points:

• BioREx achieved a new state-of-the-art F-1 score of 79.6% on the BioRED corpus, outperforming benchmark systems trained on individual datasets.
• The combined dataset improved performance across five different RE tasks.
• BioREx demonstrated robustness and generalizability on unseen tasks, including drug-drug N-ary combination and document-level gene-disease RE.

Conclusions:

• The proposed framework and BioREx approach effectively overcome limitations of existing RE methods by leveraging combined datasets.
• This work advances biomedical NLP by providing a scalable and high-performing solution for relation extraction.
• The developed tool and dataset offer significant potential for accelerating biomedical research and knowledge discovery.