GDReCo: Fine-grained gene-disease relationship extraction corpus

Area of Science:

• Biomedical Informatics
• Computational Biology
• Natural Language Processing

Background:

• Understanding gene-disease relationships is vital for medical research and drug discovery.
• Existing Natural Language Processing (NLP) models lack high-quality, fine-grained training data for knowledge extraction.

Purpose of the Study:

• To address the absence of a formal descriptive system and training corpus for NLP models in gene-disease association studies.
• To develop a novel ontology framework and a comprehensive dataset for gene-disease relationship extraction.

Main Methods:

• Development of a novel ontology framework for gene-disease associations.
• Creation of the Gene Disease Relationship Extraction Corpus (GDReCo), a dataset comprising over 24,000 instances.
• Manual annotation of 2300+ instances and model-based prediction for over 22,000 instances.

Main Results:

• BERT-based NLP models trained on GDReCo achieved high F1-scores for event and relation extraction.
• The developed corpus demonstrates effectiveness for Gene-Disease Relationship Extraction (GDRE) tasks.
• Validation of the corpus's utility through high performance metrics in relation extraction.

Conclusions:

• GDReCo is a valuable resource for advancing biomedical research and knowledge extraction.
• The study highlights the potential of NLP models for analyzing gene-disease relationships.
• Limitations of current models like ChatGPT in fine-grained relation extraction were identified.