Discovering drug-target interaction knowledge from biomedical literature

Area of Science:

• Biomedical Informatics
• Computational Biology
• Drug Discovery

Background:

• Drug-target interactions (DTI) are vital for biomedical research and drug development.
• The rapid growth of biomedical literature necessitates automated methods for DTI knowledge discovery.
• Existing extractive methods for DTI discovery are annotation-intensive and costly.

Purpose of the Study:

• To develop an efficient, end-to-end generative approach for discovering drug-target interactions (DTI) from biomedical literature.
• To overcome the limitations of traditional extractive methods that require extensive manual annotations.
• To introduce a semi-supervised learning strategy to leverage unlabeled literature for DTI discovery.

Main Methods:

• Utilized a Transformer-based generative model to directly predict DTI triplets as sequences.
• Developed a semi-supervised method to filter and label unlabeled biomedical literature.
• Trained and evaluated the model on a newly created dataset, KD-DTI.

Main Results:

• The generative approach significantly outperformed existing extractive baselines in DTI discovery.
• The proposed semi-supervised method effectively utilized unlabeled data to enhance DTI knowledge extraction.
• A new dataset, KD-DTI, was created and released to facilitate future research in this area.

Conclusions:

• Generative models offer a powerful and efficient alternative for automated DTI discovery.
• The developed semi-supervised method enhances the scalability of DTI knowledge extraction from large-scale biomedical text.
• The KD-DTI dataset and open-source code will accelerate advancements in DTI research.