Drug knowledge discovery via multi-task learning and pre-trained models.

Area of Science:

• Biomedical Natural Language Processing
• Computational Biology
• Bioinformatics

Background:

• Drug repurposing identifies new therapeutic uses for existing drugs.
• The Active Gene Annotation Corpus (AGAC) supports knowledge discovery for drug repurposing.
• The AGAC track presents challenges in selective annotation for sequence labeling tasks.

Purpose of the Study:

• To develop methods for trigger word detection (Task 1) and thematic role identification (Task 2) within the AGAC track.
• To advance drug repurposing research through automated medical knowledge extraction.
• To improve the collection of mutation-disease knowledge from biomedical literature.

Main Methods:

• Task 1 treated as Named Entity Recognition (NER) for gene mutations; Task 2 as relation extraction for entity roles.
• Utilized pre-trained biomedical language models (BioBERT, BERT, NCBI BERT, ClinicalBERT) within an information extraction pipeline.
• Employed a multi-task learning framework with fine-tuning and additional features for enhanced performance.
• Investigated knowledge consolidation and transfer from diverse sources, including ensemble integration of models.

Main Results:

• Achieved first place in Task 1 with the highest Precision (0.63), Recall (0.56), and F-score (0.60), outperforming the baseline by 0.10.
• Secured a high F-score (0.25) in Task 2 using a simple yet effective framework.
• Demonstrated the effectiveness of shared encoding layers for both NER and relation extraction tasks.

Conclusions:

• Integrating pre-trained biomedical language models with multi-task learning enhances mutation-disease knowledge extraction from PubMed.
• The developed methods show significant improvements in identifying gene mutations and their functional roles.
• The approach contributes to large-scale automatic extraction of medical knowledge, supporting drug repurposing initiatives.