Relation-aware multimodal knowledge graph learning for drug–drug interaction prediction

Drug–drug interactions (DDIs) can alter therapeutic efficacy and lead to adverse events, making accurate DDI prediction important for drug safety assessment. Existing computational methods often rely on limited drug features or do not sufficiently capture heterogeneous biomedical relations surrounding drugs. We present RA-DDI, a relation-aware knowledge graph framework for DDI prediction. RA-DDI constructs a drug-centered knowledge graph (KG) from DrugBank and initializes drug representations by integrating molecular structures, textual descriptions, and graph-derived relational features. These multimodal representations are then refined through relation-aware graph learning to capture multi-hop dependencies among drugs and related biomedical entities. A prediction module combines the learned drug embeddings with pairwise interaction features to classify DDI types. We evaluated RA-DDI on three public benchmark datasets, DrugBank, DDInter, and PDD-Graph. The proposed framework achieved competitive performance and showed consistent improvements over representative baselines on several key classification and ranking metrics. These findings demonstrate that combining multimodal drug representation learning with relation-aware knowledge graph modeling is effective for DDI prediction.