A Novel Graph Transformer Framework for Predicting Drug-Disease Associations with Structural Awareness

Area of Science:

• Computational biology
• Bioinformatics
• Drug discovery

Background:

• Accurate drug-disease association (DDA) prediction accelerates novel therapeutic discovery.
• Graph representation learning models are efficient but limited by structural biases and inability to capture complex molecular contexts.
• Existing methods struggle to learn expressive drug and disease representations from heterogeneous biomedical data.

Purpose of the Study:

• To propose SGTL-DDA, a novel graph transformer framework for enhanced DDA prediction.
• To integrate structural information and domain-specific knowledge from heterogeneous biological information networks (HBINs).
• To overcome limitations of existing models in capturing rich molecular contexts and learning expressive representations.

Main Methods:

• Developed SGTL-DDA, a graph transformer framework incorporating structural information and domain knowledge from HBINs.
• Integrated a meta-path-guided sampling strategy with a multi-level attention mechanism.
• Enabled joint learning of structural dependencies and attribute semantics in an end-to-end manner.

Main Results:

• SGTL-DDA consistently outperformed state-of-the-art methods on two benchmark datasets.
• Achieved superior performance in Accuracy, F1-score, and AUC using ten-fold cross-validation.
• Case studies on Alzheimer's disease and breast cancer validated predictive capabilities, identifying known and novel drug candidates.

Conclusions:

• SGTL-DDA effectively predicts drug-disease associations by integrating structural and semantic information from HBINs.
• The model demonstrates significant potential for accelerating drug discovery and repositioning.
• Validated findings through molecular docking and literature evidence, confirming predictive accuracy and novel candidate identification.