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

A bond-level substructure-selective encoder is introduced in molecular line graphs, performing selective information routing between adjacent bond states and condensing interaction-relevant substructures into compact molecular representations. Relative to existing molecular graph learning methods that rely on atom-centered propagation and global graph aggregation, the design shifts modeling emphasis to bond-level local structure so that interaction-relevant local patterns are emphasized. The abstract reports outperforming representative baselines across multiple metrics on DrugBank and Twosides under transductive and inductive settings, with ablation and visualization analyses supporting selective encoding; specific values are not given in the text.

Beyond local structural modeling, SMILES-derived Morgan-fingerprint similarity profiles are refined to suppress noisy global structural signals. This refinement does not rely on external biomedical annotations, offering a complementary global-signal processing path for local structural representations. Stated in the abstract as a component of the framework, with its separate contribution indirectly supported by ablation analysis; the text provides no isolated quantitative result.

The overall SSE-DDI framework targets relation-specific DDI prediction and reports results under both transductive and inductive evaluation settings. Covering both transductive and inductive settings means evaluation is not limited to cases of known drug combinations. Based on experiments on the DrugBank and Twosides datasets, the abstract states outperformance over representative baselines across multiple metrics; specific metrics and effect sizes are not listed in the text.

The work targets relation-specific DDI prediction with molecular structure as input, applies to the transductive and inductive evaluation settings represented by DrugBank and Twosides, and explicitly avoids external biomedical annotations, making it suited to early screening and medication-safety pre-screening where only molecular structure information is available; its design goal is to offer modelers attentive to local substructure signals a bond-centered representation path.

The text is abstract-level information and does not list specific metric values, dataset sizes, baseline lists, or ablation details, so the magnitude and stability of the performance gains are hard to judge; how bond-level selective encoding behaves across different chemical spaces, different interaction types, and real-world clinical polypharmacy populations, and under which conditions the refined fingerprint similarity profiles are most beneficial, remain questions worth watching.