Predicting biomedical relationships using the knowledge and graph embedding cascade model.

Area of Science:

• Biomedical informatics
• Machine learning in drug discovery
• Network biology

Background:

• Large-scale pharmacological, genomic, and chemical datasets enable new insights into biochemical networks.
• Existing knowledge embedding models struggle with unbalanced and sparse biochemical networks.
• Graph embedding offers complementary features to address limitations of knowledge embedding.

Purpose of the Study:

• To develop a hybrid embedding approach for representing biochemical entities and their relationships.
• To create a cascade learning framework integrating knowledge and graph embeddings for link prediction.
• To evaluate the proposed model's performance against existing methods in a biomedical context.

Main Methods:

• Combined knowledge embedding and graph embedding to create dense, low-dimensional vector representations.
• Developed a cascade learning framework incorporating semantic and graph features for link probability scoring.
• Designed a meta-path algorithm for detecting complex path relations within biomedical networks.

Main Results:

• Achieved 93% accuracy in predicting links and entities.
• Demonstrated an average 8.6% absolute improvement in hits@10 score compared to standalone knowledge embedding.
• Showed 1.1% to 9.7% absolute improvement over other knowledge and graph embedding algorithms.

Conclusions:

• The proposed hybrid embedding model effectively represents biochemical networks and enhances link prediction.
• The cascade learning framework offers a robust solution for analyzing sparse and unbalanced biological data.
• Case studies, including a potential link between vitamin D receptor (VDR) and prostate cancer, highlight the model's utility for biomedical researchers.