SSE-DDI: Selective Substructure Encoding with Bond-Centered Molecular Representations for Drug-Drug Interaction Prediction.

Drug-drug interactions (DDIs) can alter therapeutic efficacy or cause severe adverse reactions, posing major risks in polypharmacy. Accurate DDI prediction is therefore important for medication safety and early stage drug screening. Although molecular graph learning has been widely used for DDI prediction, many methods still rely on atom-centered propagation and global graph aggregation, which may obscure interaction-relevant local structural patterns. To address this limitation, we propose SSE-DDI, a molecular structure-driven framework centered on selective substructure encoding for relation-specific DDI prediction. SSE introduces a bond-level substructure-selective encoder in molecular line graphs, where chemical bonds serve as fundamental representation units and DDI-relevant local patterns are emphasized before drug-level representation learning. The encoder performs selective information routing between adjacent bond states and condenses interaction-relevant substructures into compact molecular representations. These representations are further refined by an edge-fusion graph transformer. To complement local structural modeling, SSE-DDI refines SMILES-derived Morgan-fingerprint similarity profiles to suppress noisy global structural signals without external biomedical annotations. Experiments on DrugBank and Twosides under transductive and inductive settings show that SSE-DDI outperforms representative baselines across multiple metrics. Ablation and visualization analyses verify the effectiveness of selective encoding and highlight DDI-relevant molecular substructures. The source code is available at https://github.com/qichi77/SSE-DDI.