PT-KGNN: A framework for pre-training biomedical knowledge graphs with graph neural networks.

Area of Science:

• Bioinformatics
• Computational Biology
• Artificial Intelligence in Medicine

Background:

• Biomedical knowledge graphs (KGs) are crucial for modeling complex biological relationships.
• Existing methods for feature learning in KGs often use traditional machine learning or graph neural networks (GNNs).
• Pre-training techniques from natural language processing (NLP) offer a promising avenue for enhancing KG representations.

Purpose of the Study:

• To propose and evaluate a novel framework, PT-KGNN, for pre-training biomedical KGs using GNNs.
• To investigate the impact of KG scale on the effectiveness of the pre-training approach.
• To enhance the performance of downstream tasks such as drug-drug interaction (DDI) and drug-disease association (DDA) prediction.

Main Methods:

• Developed the PT-KGNN framework, which applies GNNs to biomedical KGs for node embedding learning.
• Conducted experiments to assess the framework's effectiveness and the influence of KG scale.
• Evaluated performance on independent datasets for DDI and DDA prediction tasks.

Main Results:

• The PT-KGNN framework demonstrated improved node embeddings through pre-training.
• Performance consistently improved with the increasing scale of the biomedical KG used for pre-training.
• Pre-training on large-scale KGs significantly boosted DDI and DDA prediction accuracy.

Conclusions:

• Pre-training biomedical KGs with GNNs effectively captures rich semantic and structural information.
• The scale of the KG is a critical factor, with larger KGs yielding superior results.
• This approach holds significant potential for advancing various bioinformatics applications.