Improving target-disease association prediction through a graph neural network with credibility information

Area of Science:

• Bioinformatics
• Computational Biology
• Drug Discovery

Background:

• Identifying target-disease associations (TDAs) is crucial for drug development, but current methods face challenges with data scarcity and credibility.
• Existing machine learning models for TDAs often lack guaranteed credibility and struggle with predicting associations for underrepresented entities.

Purpose of the Study:

• To develop an effective deep learning framework, CreaTDA, for predicting novel target-disease associations (TDAs).
• To enhance TDA prediction by incorporating credibility information from literature and improving performance on sparse biological networks.

Main Methods:

• Utilized graph neural networks for feature extraction from heterogeneous biological data.
• Developed an end-to-end deep learning framework (CreaTDA) to learn latent representations of targets and diseases.
• Integrated literature-derived credibility information to improve TDA prediction accuracy.

Main Results:

• CreaTDA demonstrated superior prediction performance compared to state-of-the-art methods on comprehensive and sparse TDA networks.
• The framework successfully predicted novel TDAs with supporting evidence from existing literature.
• Achieved significant improvements in predicting associations for proteins with limited known disease connections.

Conclusions:

• CreaTDA provides a robust tool for identifying novel target-disease associations, significantly aiding the drug discovery process.
• The integration of credibility information enhances the reliability and novelty of predicted TDAs.
• The framework shows promise in overcoming limitations of current models, particularly for data-scarce entities.