BioGraphFusion: graph knowledge embedding for biological completion and reasoning.

Area of Science:

• Biomedical informatics
• Computational biology
• Artificial intelligence in medicine

Background:

• Biomedical knowledge graphs (KGs) are vital for drug discovery and disease understanding.
• Existing methods like knowledge embedding (KE) and graph neural networks (GNNs) have limitations in capturing both global semantics and local structures.
• A synergistic approach is needed for adaptive learning in complex biomedical KGs.

Purpose of the Study:

• To introduce BioGraphFusion, a novel framework for deep synergistic semantic and structural learning in biomedical KGs.
• To address the limitations of current methods in achieving adaptive interplay between semantic and structural learning.
• To improve the completion and reasoning capabilities of biomedical KGs.

Main Methods:

• BioGraphFusion utilizes tensor decomposition for a global semantic foundation.
• An LSTM-driven mechanism dynamically refines relation embeddings during graph propagation.
• Query-guided subgraph construction and a hybrid scoring mechanism enhance adaptive learning.

Main Results:

• BioGraphFusion demonstrated superior performance over state-of-the-art KE, GNN, and ensemble models.
• The framework achieved superior results across three key biomedical tasks.
• A case study on cutaneous malignant melanoma revealed biologically meaningful pathways.

Conclusions:

• BioGraphFusion offers a novel and effective approach for synergistic semantic and structural learning in biomedical KGs.
• The framework shows significant potential for advancing drug discovery and disease understanding.
• Open-source code is available for reproducibility and further development.