MvDDI: A Multi-view Interaction Framework for Few-Shot Drug-Drug Interaction

Zihao Mao, Qiguang Chen, Yongheng Zhang, Ruoxi Zhou, Peng Wang, Yao Li, Sheng Wang, Libo Qin

Abstract: Drug-drug interaction (DDI) poses significant risks to patient safety and has garnered increasing attention. In recent years, advances in deep learning and pre-trained models have driven substantial progress in DDI prediction. While these models demonstrate strong predictive performance, their effectiveness heavily relies on the availability of large labeled datasets, which limits their generalization in real-world scenarios. To address this challenge, we propose a novel task: few-shot DDI prediction, aimed at mitigating the issue of data scarcity. We also introduce four benchmarks under both transductive and inductive settings—Few-Drugbank, Few-TWOSIDES, Few-S1, and Few-S2—to facilitate further research. To tackle the few-shot DDI problem, we propose the Multi-view Drug-Drug Interaction framework (MvDDI), which integrates information from three distinct views. Experimental results demonstrate that MvDDI achieves state-of-the-art performance in few-shot DDI prediction and outperforms existing models on full-shot datasets, delivering the best results.

External IDs:dblp:conf/nlpcc/MaoCZZWLWQ25