MRDGNN: A multi-relational reasoning framework for predicting drug indications via relational digraphs

Area of Science:

• Biomedical Informatics
• Computational Biology
• Drug Discovery

Background:

• Predicting drug indications is crucial for drug repurposing and biomedical research.
• Clinically relevant signals, including contraindications, offer complementary evidence for novel indication discovery.
• Existing methods may not fully leverage complex relational data for drug-disease association prediction.

Purpose of the Study:

• To develop a novel framework, MRDGNN (Multi-Relational Digraph-based Graph Neural Network), for enhanced drug-disease association prediction.
• To incorporate multi-hop reasoning over relational digraphs and distinguish between opposite drug-disease relationships.
• To leverage auxiliary biomedical knowledge through a pretraining-fine-tuning strategy.

Main Methods:

• MRDGNN employs recursively constructed relational digraphs (r-digraphs) for multi-hop reasoning.
• A query-conditioned relation-aware attention mechanism distinguishes opposite drug-disease relations.
• Layer-wise attention adaptively integrates evidence from varying reasoning depths.
• A pretraining-fine-tuning strategy on PrimeKG, excluding direct drug-disease associations, provides biologically informed initialization.

Main Results:

• MRDGNN consistently outperformed competitive baselines in drug indication prediction on the PrimeKG dataset, based on discrimination and ranking metrics.
• The model demonstrated reasonable effectiveness in held-out-drug cold-start and sparse-relation settings.
• Performance was positively correlated with the availability of relational evidence.

Conclusions:

• MRDGNN offers a powerful framework for predicting drug indications and facilitating drug repurposing.
• The model's ability to handle complex relational data and distinguish opposite signals enhances prediction accuracy.
• A case study on Alzheimer's disease highlighted MRDGNN's utility in identifying candidate drugs with interpretable evidence.