Abstract
Drug repositioning has become attractive because it can significantly accelerate drug discovery and reduce development costs by identifying new indications for existing drugs. Based on graph neural networks (GNNs), the recent work utilizes inter-domain information (drug-disease association network) and intra-domain information (drug-drug similarity network and disease-disease similarity network) to learn the effective representation of drugs and diseases. However, they overlook the significant impact of the adaptive inter-domain and intra-domain information fusion operation on model performance under different data-splitting strategies. Moreover, manually designing GNN architectures for specific drug repositioning datasets is time-consuming and expert-dependent. To address the above problem, we propose an adaptive drug repositioning prediction method called AdaDR, which can adaptively fuse inter-domain and intra-domain information under different data-splitting strategies and automatically design the optimal GNN architecture for each drug repositioning dataset. Specifically, we first design a unified drug repositioning search space with different information fusion operations and various handcrafted GNN architectures. Then, a drug repositioning model search will be adopted to enable an efficient search. Empirical studies on three benchmark datasets demonstrate that the optimal drug repositioning model identified by our proposed AdaDR achieves the best performance among competitive baselines. Through the analysis of the case study, the applicability of AdaDR in practical scenarios is further validated. The code is available at: https://github.com/csubigdata-Organization/AdaDR.