Multimodal reasoning based on knowledge graph embedding for specific diseases.

Area of Science:

• Biomedical informatics
• Knowledge representation
• Artificial intelligence in medicine

Background:

• Biomedical knowledge graphs (KGs) are crucial for extracting new information.
• Existing KGs often lack disease-specific focus.
• Multimodal reasoning enhances KG-based knowledge discovery.

Purpose of the Study:

• To develop a construction and multimodal reasoning framework for Specific Disease Knowledge Graphs (SDKGs).
• To create SDKG-11, a dataset covering five cancers and six non-cancer diseases.
• To enable discovery of new, reliable disease-specific knowledge and provide universal pre-trained models.

Main Methods:

• Constructed SDKG-11 via triplet extraction, entity/relation linking.
• Implemented multimodal reasoning using reverse-hyperplane projection on structure, category, and description embeddings.
• Evaluated performance using entity prediction and manual verification of novel drug-gene, gene-disease, and disease-drug pairs.

Main Results:

• SDKG-11 construction completed, encompassing diverse diseases.
• Multimodal reasoning significantly improved pre-existing models across all SDKGs.
• Validated the discovery of novel, reliable biomedical knowledge pairs.
• Demonstrated the universality of embedding models for biomolecular interaction classification.

Conclusions:

• SDKGs offer a powerful approach for disease-specific knowledge discovery.
• Multimodal reasoning enhances the utility of biomedical KGs.
• The developed SDKG-11 dataset and framework provide a valuable resource for researchers.