Subgraph-based methods have demonstrated strong and interpretable capabilities in predicting drug-drug interactions (DDIs), which are crucial for both medical practice and drug development.Subgraph selection and encoding are critical stages in the framework of these methods. Customizing these components is promising but relatively unexplored due to the high cost of manual tweaking.In this work, inspired by the success of neural architecture search (NAS), we propose to search for data-specific components of the subgraph-based pipeline.In particular, we introduce comprehensive subgraph selection and encoding spaces to cover the diverse contexts of drug interactions for DDI prediction.Faced with vast spaces and overwhelming sampling overhead, we design an effective relaxation mechanism to efficiently explore optimal subgraph configurations using an approximation strategy, enabling a robust search algorithm to explore the search space efficiently.Extensive experiments illustrate the effectiveness and superiority of the proposed method. Additionally, the searched subgraphs and encoding functions clearly demonstrate the model's adaptivity.