IS-DDI: Invariant Substructure Based Method for Drug-Drug Interaction Prediction

Combination medication therapy has grown in popularity in the treatment of diseases; nevertheless, it also increases the possibility of drug-drug interactions (DDI), which can have unfavorable consequences on the body. As such, precise DDI prediction is vitally important. Several deep-learning approaches based on drug substructures have been proposed in recent years to predict DDI. However, these methods often struggle to generalize to unfamiliar domains, and their approaches to substructure extraction can be suboptimal. In this paper, we propose IS-DDI, a novel DDI prediction model based on invariant substructures. We employ a domain-invariant loss function to enhance generalization ability and adopt a Transformer-like substructure extraction module to obtain self-adaptive substructures effectively. We evaluate IS-DDI in both transductive and inductive setting. Experimental results demonstrate that IS-DDI not only has the advantage in prediction performance but also exhibits robust cross-domain generalization ability.