A geometric deep learning framework for drug repositioning over heterogeneous information networks

Area of Science:

• Biomedical Informatics
• Computational Drug Discovery
• Artificial Intelligence in Medicine

Background:

• Drug repositioning (DR) accelerates drug discovery but often neglects the non-Euclidean nature of biomedical network data.
• Existing computational methods for DR may not fully capture complex biological relationships.
• The integration of artificial intelligence (AI) offers a promising avenue for enhancing traditional drug discovery and development.

Purpose of the Study:

• To propose a novel deep learning framework, DDAGDL, for predicting drug-drug associations (DDAs).
• To address the limitations of current DR methods by incorporating the non-Euclidean characteristics of biomedical networks.
• To improve the efficacy of drug repositioning through advanced computational techniques.

Main Methods:

• Development of DDAGDL, a deep learning framework utilizing geometric deep learning (GDL) on heterogeneous information networks (HIN).
• Incorporation of complex biological information into the topological structure of HIN.
• Application of an attention mechanism for learning smoothed representations of drugs and diseases.

Main Results:

• DDAGDL demonstrated superior performance on three real-world datasets compared to state-of-the-art DR methods.
• The framework achieved high efficacy across various evaluation metrics under 10-fold cross-validation.
• Case studies and molecular docking experiments validated DDAGDL's potential as a promising DR tool.

Conclusions:

• DDAGDL effectively predicts drug-drug associations by leveraging geometric deep learning on heterogeneous information networks.
• The framework offers a significant advancement in drug repositioning by integrating complex biological data and network topology.
• Exploiting geometric prior knowledge with DDAGDL provides new insights for improved drug efficacy and discovery.