Chemical-protein interaction extraction via Gaussian probability distribution and external biomedical knowledge

Area of Science:

• Biomedical Informatics
• Computational Biology
• Natural Language Processing

Background:

• Biomedical literature contains extensive chemical-protein interaction (CPI) data crucial for drug discovery and research.
• Current CPI extraction methods often overlook local sentence structure and external biomedical knowledge, limiting performance.

Purpose of the Study:

• To develop a novel neural network approach for improved chemical-protein interaction (CPI) extraction.
• To leverage local sentence structure and external biomedical knowledge to enhance CPI identification accuracy.

Main Methods:

• Utilized BERT for contextual representations of title, instance, and knowledge sequences.
• Incorporated Gaussian probability distribution to model local instance structure.
• Applied attention mechanisms to integrate title and biomedical knowledge.
• Concatenated representations and used softmax for CPI extraction.

Main Results:

• The proposed model demonstrated superior performance on the CHEMPROT corpus compared to state-of-the-art methods.
• Gaussian probability distribution and external knowledge were found to be complementary, significantly improving CPI extraction.
• The Gaussian probability distribution enhanced extraction for sentences with overlapping relations.

Conclusions:

• The novel neural network approach effectively improves chemical-protein interaction (CPI) extraction by integrating local structure and external knowledge.
• Combining Gaussian probability distribution and external knowledge offers a synergistic improvement in identifying CPIs.
• This method shows promise for advancing drug discovery and biomedical research through more accurate information extraction.