DMSDR: Drug Molecule Synergy-Enhanced Network for Drug Recommendation with Multi-source Domain Knowledge

The drug recommendation provides safe drug combinations based on the patient’s medical conditions. Existing research efforts have focused on recommending drugs by modeling patient electronic health records and drug side effects knowledge. However, they overlook the importance of drug synergy in drug recommendation, despite the fact that synergistic drug combinations can enhance therapeutic efficacy. To address this gap, we propose the Drug Molecule Synergy-Enhanced Network for Drug Recommendation with Multi-Source Domain Knowledge (DMSDR), which incorporates the concept of drug synergy for the first time. Specifically, we design a drug molecular synergy module to extract drug features at the molecular structure level and learn the interaction information between different drug pairs to identify truly effective drug combinations for patients. Furthermore, we introduce medical external knowledge through a drug synergy interaction graph and model the synergistic effects between drugs using graph convolutional neural networks. To ensure the safety of recommended drugs, we incorporate a contrastive loss function during training to learn both drug synergy and drug interaction features, thereby minimizing potential drug side effects. Extensive experiments conduct on the MIMIC-III and MIMIC-IV datasets demonstrate that our model outperforms multiple baselines and achieves significant performance improvements. The source code is available at: https://github.com/AK-321/DMSDR .