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

Predicting drug indications is a fundamental task in biomedical research and drug repurposing. In addition to known therapeutic associations, clinically relevant but opposite signals, such as contraindications, may provide complementary evidence for identifying novel indications. In this work, we propose MRDGNN (Multi-Relational Digraph-based Graph Neural Network), a framework for drug-disease association prediction that performs multi-hop reasoning over recursively constructed relational digraphs (r-digraphs). MRDGNN incorporates a query-conditioned relation-aware attention mechanism to distinguish clinically opposite drug-disease relations and employs layer-wise attention to adaptively integrate evidence from different reasoning depths. To provide biologically informed initialization, we further adopt a pretraining-fine-tuning strategy on PrimeKG, where drug-disease associations are excluded during pretraining to avoid direct information leakage while preserving rich auxiliary biomedical knowledge. Experiments on PrimeKG show that MRDGNN consistently outperforms competitive baselines on indication prediction across both discrimination and ranking metrics. Additional evaluations under held-out-drug cold-start and sparse-relation settings further show that the model remains reasonably effective under less favorable conditions, although its performance is still positively associated with the availability of relational evidence. A case study on Alzheimer's disease further demonstrates the utility of MRDGNN by identifying plausible candidate drugs and providing interpretable relational evidence for its predictions.