DUVEL: an active-learning annotated biomedical corpus for the recognition of oligogenic combinations

Area of Science:

• Bioinformatics
• Computational Biology
• Genomics

Background:

• Biomedical relation extraction (bioRE) datasets currently focus on single variants, limiting research into complex genetic interactions.
• Understanding digenic and oligogenic variant combinations is vital for elucidating disease etiologies and epistatic effects.
• Existing literature highlights the importance of multi-variant interactions, yet lacks dedicated datasets for computational analysis.

Purpose of the Study:

• To create a unique dataset (DUVEL) for training computational tools to extract oligogenic variant combinations from scientific literature.
• To address the scarcity of datasets for multi-gene variant interactions in biomedical relation extraction.
• To leverage active learning to optimize the annotation process for efficiency and cost-effectiveness.

Main Methods:

• Utilized active learning (AL) to guide the annotation of text fragments containing potential digenic variant combinations (gene-variant-gene-variant).
• Pre-annotated 85 full-text articles from the Oligogenic Diseases Database (OLIDA) using PubTator.
• Annotated extracted text fragments using ALAMBIC, an AL-based annotation platform, to create the DUVEL dataset.

Main Results:

• The DUVEL dataset comprises 8442 text fragments, with 794 positive instances of oligogenic variant relations, covering 95% of annotated articles.
• Fine-tuning state-of-the-art biomedical language models (BiomedBERT, BiomedBERT-large, BioLinkBERT, BioM-BERT) on DUVEL significantly improved performance.
• BiomedBERT-large achieved the highest F1 score of 0.84 for gene-variant pair detection after fine-tuning, demonstrating the dataset's utility.

Conclusions:

• The DUVEL dataset provides a valuable resource for advancing biomedical relation extraction, specifically for 4-ary relations between two genes and two variants.
• Active learning is an effective strategy for creating specialized bioRE datasets, reducing annotation costs and improving efficiency.
• The DUVEL dataset is freely available, facilitating research in computational biology and the curation of scientific literature on complex genetic interactions.