A Task‐Aware Framework for Asymmetric Drug–Drug Interaction Prediction

ABSTRACT The prediction of Drug–Drug Interactions (DDIs) is critical for patient safety and therapeutic efficacy, given the potential for unknown physicochemical incompatibilities between co‐administered drugs. However, while existing methods can effectively capture the chemical information of drugs, they generally overlook the inherent asymmetry of drug interactions. Furthermore, most current prediction models are built on a transductive learning paradigm, which lacks adaptability to inductive prediction scenarios involving novel drugs. This significantly limits the generalization capability and practical applicability of these models in real‐world settings. To address these challenges, we introduce a novel framework named Task‐adaptive Dual‐role learning for DDI (TaRo‐DDI), which effectively captures the directional asymmetric information of DDIs while maintaining high generalization capability, thereby enhancing prediction accuracy. First, the model extracts Morgan fingerprints from the chemical structures of drugs and constructs a directed DDI network, thereby providing rich contextual information for DDI prediction. Second, depending on whether the prediction task involves new drugs, the framework dynamically orchestrates distinct specialized engines. For transductive tasks, a relation decoding engine is designed to learn differentiated role‐specific features by modeling complex relationships between known drug pairs; for inductive tasks, a representation learning engine is developed to effectively extract distinct role features for both new drugs and known drugs, even in the absence of historical interaction data. Finally, an attention mechanism is introduced to weight and fuse multimodal features adaptively through a deep neural network, further enhancing the model's ability to capture complex drug relationships and its generalization performance. Experiments demonstrate that TaRo‐DDI outperforms or matches all baseline methods across all evaluation metrics in both transductive and inductive tasks. Specifically, in transductive tasks, it achieves improvements of 0.3% in AUROC and 0.6% in AUPRC over the best baseline. In the more challenging inductive tasks, the improvements are 0.1% in AUROC and 0.5% in AUPRC over the respective best baseline in each metric.