An inductive learning-based method for predicting drug-gene interactions using a multi-relational drug-disease-gene...

Area of Science:

• Computational biology
• Bioinformatics
• Machine learning

Background:

• Drug-gene interactions (DGIs) are crucial for drug discovery.
• Current transductive learning models struggle with predicting interactions involving entirely new drugs or genes (unseen DGIs).
• Data sparsity is a challenge in DGI prediction.

Purpose of the Study:

• To develop an inductive learning-based model for precise identification of unseen drug-gene interactions (DGIs).
• To improve the prediction of both known and novel DGIs.

Main Methods:

• Constructed a multi-relational drug-disease-gene (DDG) graph integrating disease nodes to mitigate data sparsity.
• Extracted graph features using graph embedding algorithms and retrieved individual gene/drug node attributes.
• Developed a hybrid feature representation by combining graph features and node attributes.
• Implemented an innovative inductive learning approach by transforming known node vectors into unseen node representations using node similarities as weights.

Main Results:

• The proposed inductive learning model significantly outperformed existing models in predicting both external unknown and unseen DGIs.
• The model demonstrated practical feasibility through case studies and molecular docking.
• Achieved accurate and cost-effective DGI detection.

Conclusions:

• This study presents an efficient, data-driven approach for DGI prediction using inductive learning.
• The model offers a promising tool for accelerating drug discovery and repurposing by identifying novel interactions.
• The integration of disease nodes and hybrid feature characterization enhances prediction accuracy.