Cross-modal embedding integrator for disease-gene/protein association prediction using a multi-head attention...

Area of Science:

• Biomedical informatics
• Computational biology
• Machine learning in healthcare

Background:

• Biomedical knowledge graphs are crucial for inferring new knowledge and identifying disease-gene/protein relationships.
• Accurate representations of biomedical entities are vital for predicting these associations.
• Discovering novel disease-gene/protein links can uncover disease mechanisms and therapeutic targets.

Purpose of the Study:

• To develop a computational model for predicting disease-gene/protein associations.
• To integrate diverse data modalities for improved prediction accuracy.
• To leverage biomedical knowledge graphs for enhanced biological insights.

Main Methods:

• Utilized the Precision Medicine Knowledge Graph.
• Generated biomedical entity embeddings using a large language model (LLM) and a knowledge graph embedding (KGE) algorithm.
• Developed the Cross-Modal Embedding Integrator (CMEI) model, integrating embeddings via multi-head attention.

Main Results:

• The CMEI model achieved a high predictive performance with an area under the receiver operating characteristic curve of 0.9662 (±0.0002).
• Demonstrated the effectiveness of integrating LLM and KGE embeddings for association prediction.
• Validated the model's capability in predicting disease-gene/protein relationships.

Conclusions:

• The developed computational model, CMEI, effectively predicts disease-gene/protein associations.
• CMEI shows potential for accelerating the identification of disease development mechanisms.
• This approach may contribute to discovering new therapeutic targets for diseases.