Adaptive debiasing learning for drug repositioning

Drug repositioning, pivotal in current pharmaceutical development, aims to find new uses for existing drugs, offering an efficient and cost-effective path to drug discovery. In recent years, graph neural network-based deep learning methods have achieved significant success in drug repositioning tasks. However, few studies have analyzed the characteristics of datasets to mitigate potential data biases. In this paper, we analyzed three commonly used drug repositioning datasets and identified a consistent characteristic among them: a trend of node polarization, characterized by the presence of popular entities (those commonly occurring and extensively associated) and long-tail entities (those appearing less frequently with fewer associations). Based on this finding, we propose a deep learning framework with a debiasing mechanism, called DRDM. The framework excels in addressing popular entities’ biases, which often overshadow the subtle patterns in long-tail entities—key for novel insights. DRDM dynamically adjusts association weights during training, enhancing long-tail entity representation and reducing bias. In addition, we employ dual-view contrastive learning to provide rich supervisory signals, thereby further enhancing the model’s robustness. We conducted experiments with our method on these three datasets, and the results demonstrated that our approach exhibits strong competitiveness compared to competing models. Case studies further highlighted the potential of the model in practical applications, which could provide new insights for future drug discovery.