Predicting Drug-Gene Interaction via Multi-Negative Sampling-Based BPR Contrastive Learning

Graph Neural Networks have recently demonstrated significant potential in enhancing drug-gene interaction prediction. GNN-based methods have emerged as a promising approach due to their robust capability to model connections within the drug-gene bipartite graph. Existing approaches can be broadly classified into two categories: subgraph extraction methods based on random walks and contrastive learning methods relying on dual views. Subgraph extraction methods leveraging random walks effectively utilize local information for drug-gene interaction prediction; however, their performance is constrained by the limited scope of local data. Conversely, contrastive learning methods, which depend on comparing two fixed views, struggle to capture the complexity and diversity of drug-gene interaction, thus restricting the model’s ability to discern finer interaction patterns. Moreover, in real-world drug discovery scenarios, interaction data are often imbalanced, introducing model bias that compromises both prediction accuracy and practical applicability. To address these limitations, we propose a novel Multi-Negative Sample Contrastive Learning (MNCL) framework based on Bayesian Personalized Ranking (BPR). This framework adopts BPR as its core task and integrates both global and local negative sampling strategies, significantly enhancing model performance and improving its capacity to predict diverse relationship types. Additionally, it mitigates bias arising from label imbalance. Extensive experiments show that MNCL outperforms state-of-the-art baseline methods on the DrugBank and DGIdb datasets, achieving accuracy improvements of 2.0% on DrugBank and 1.3% on DGIdb, with an overall accuracy of 0.946.