Chemical-protein interaction extraction via contextualized word representations and multihead attention

Area of Science:

• Biomedical Natural Language Processing (NLP)
• Computational Biology
• Pharmacology

Background:

• Biomedical literature contains vast, unstructured data on chemical-protein interactions (CPIs).
• Automated extraction of CPIs is vital for advancing pharmacological and clinical research.
• Existing methods may not fully capture complex contextual information within scientific text.

Purpose of the Study:

• To develop and evaluate a deep neural network model for enhanced automatic extraction of CPIs.
• To investigate the efficacy of deep context representation and multihead attention mechanisms in CPI extraction.
• To improve the accuracy and efficiency of identifying CPIs from biomedical literature.

Main Methods:

• A novel deep neural model integrating deep context representation, Bidirectional Long Short-Term Memory networks (Bi-LSTMs), and a multihead attention layer.
• Deep context representation generates comprehensive sentence embeddings.
• Multihead attention focuses on salient features from Bi-LSTMs outputs.

Main Results:

• The proposed model demonstrated competitive performance on the public ChemProt corpus.
• Both deep context representation and multihead attention significantly contributed to improved CPI extraction.
• The model shows potential to outperform existing state-of-the-art methods for CPI identification.

Conclusions:

• Deep context representation and multihead attention are effective components for enhancing CPI extraction models.
• The developed deep neural network offers a promising approach for automated CPI discovery.
• This work facilitates large-scale analysis of biomedical literature for drug discovery and clinical applications.