Using BERT to identify drug-target interactions from whole PubMed

Area of Science:

• Computational biology
• Bioinformatics
• Machine learning in drug discovery

Background:

• Drug-target interactions (DTIs) are crucial for drug repurposing and understanding drug mechanisms.
• Existing databases manually curate DTIs, but likely represent a small fraction of available experimental data.
• Systematic approaches are needed to identify and extract DTIs from the vast scientific literature.

Purpose of the Study:

• To apply Bidirectional Encoder Representations from Transformers (BERT) for identifying articles containing experimentally determined DTIs.
• To develop functions for predicting the assay format used in DTI studies.
• To enhance the systematic curation of DTIs and facilitate drug discovery.

Main Methods:

• Utilized Bidirectional Encoder Representations from Transformers (BERT) to identify relevant scientific articles.
• Developed a pipeline to extract drug and protein information from identified articles.
• Incorporated functions to predict the assay format for DTI data.

Main Results:

• Identified 0.6 million articles with drug and protein information not present in public DTI databases.
• Achieved approximately 99% accuracy in identifying articles with quantitative drug-target profiles via 10-fold cross-validation.
• Attained an 88% F1 micro-score for predicting assay format, indicating potential for future refinement.

Conclusions:

• The developed BERT model and pipeline are robust for identifying previously overlooked articles containing quantitative DTIs.
• This method represents a significant advancement in machine-assisted DTI extraction and curation.
• The approach is expected to be valuable for drug mechanism discovery and repurposing efforts.