Literature mining for context-specific molecular relations using multimodal representations (COMMODAR).

Area of Science:

• Biomedical informatics
• Computational biology
• Bioinformatics

Background:

• Understanding biological systems requires contextual information on complex molecular relations.
• Manual extraction of these relations from scientific literature is time-consuming and labor-intensive.
• Existing methods often lack the comprehensive representations needed for accurate context-specific relation extraction.

Purpose of the Study:

• To develop a machine learning-based literature mining framework, COMMODAR, for extracting context-specific molecular relations.
• To enhance relation extraction by utilizing multimodal representations, combining biomedical domain knowledge and linguistic information.
• To provide a publicly available resource for molecular relation extraction.

Main Methods:

• Developed COMMODAR, a framework employing feature augmentation through the cooperation of multimodal representations for relation extraction.
• Leveraged both biomedical domain knowledge and canonical linguistic information to create comprehensive textual representations.
• Applied the framework to a large corpus of 14 million PubMed abstracts.

Main Results:

• Models utilizing multiple modalities demonstrated superior performance compared to those relying solely on linguistic information.
• Successfully extracted 9214 context-specific molecular relations from the PubMed abstracts.
• The COMMODAR framework achieved high accuracy in identifying molecular relationships within their biological context.

Conclusions:

• COMMODAR offers an effective machine learning approach for automated, context-specific molecular relation extraction from biomedical literature.
• Multimodal representations are crucial for improving the accuracy and comprehensiveness of relation extraction.
• The framework and extracted data provide valuable resources for advancing biological research and knowledge discovery.