Abstract
Multidrug combination therapy has long been a vital approach for treating complex diseases by leveraging synergistic effects between drugs. However, drug-drug interactions (DDIs) are not uniformly beneficial. Accurate and rapid identification of DDIs is critical to mitigate drug-related side effects. Currently, many computational-based methods have been used to expedite the prediction of DDIs. However, most of these methods use a single perspective to obtain drug features, which have limited expressive capabilities and cannot fully represent the essential attributes of drugs. In this study, we propose the Multi-view Feature Embedding for drug-drug interaction prediction (MFE-DDI), which integrates SMILES information, molecular graph data and atom spatial semantic information to model drugs from multiple perspectives and encapsulate the intricate drug information crucial for predicting DDIs. Concurrently, the feature information extracted from different feature encoding channels is fused in the attention-based fusion module to fully convey the essence of drugs. Consequently, this approach enhances the efficacy of the DDI prediction task. Experimental results indicate that MFE-DDI surpasses other baseline methods on three datasets. Moreover, analysis experiments demonstrate the robustness of the model and the necessity of each component of the model. Case studies on newly approved drugs demonstrate the effectiveness of our method in real scenarios. The code and data used in MFE-DDI can be found at https://github.com/2019040445/MFE_DDI.