Hierarchical Structure-Aware Graph Prompting for Drug-Drug Interaction Prediction

Abstract

Drug-drug interaction (DDI) prediction holds crucial significance in biomedical applications such as polypharmacy and clinical decision-making. Considering the limited availability of labeled DDI relations, it is promising to effectively extract underlying knowledge from drug molecular graphs by self-supervised learning to enhance DDI prediction performance, owing to the recent successes in graph pre-training for molecular representation. Nonetheless, employing existing graph pretraining methods directly reveals significant disparities persisting between the pre-training tasks and the ultimate objective of DDI prediction. Addressing this, we propose HS-GPF, a novel hierarchical structure-aware graph prompting framework tailored for DDI prediction. Its key component is a specially designed graph prompt learning mechanism, which significantly integrates the pre-training and the final DDI task into a uniform task format. This is achieved through an adaptive dual-level prompting process featuring unique virtual tokens. Aligned with our hierarchical structure-aware pre-training, it effectively activates relevant knowledge for DDI prediction, fostering a more seamless integration between the pre-trained model and complex drug interactions. Extensive experiments across various scales of real-world datasets demonstrate that our method outperforms existing state-of-the-art baselines, even in challenging cold-start scenarios.