IHM-DDI: Integrating high-order interaction topology and Morgan fingerprint features for drug-drug interaction...

Drug-drug Interaction (DDI) prediction is crucial in pharmacology and clinical applications. Conventional methods remain constrained by single-view paradigms, limiting multi-source data fusion and cross-view synergistic learning critical for comprehensive drug property profiling. We introduce IHM-DDI, an integrative deep-learning framework that integrates a high-order interaction topology and Morgan fingerprint-based features to enhance DDI prediction. Specifically, IHM-DDI employs an autoencoder to reduce the dimensionality of Morgan fingerprint similarity matrices, followed by KNN-based drug similarity network graph construction to identify topologically relevant drug pairs for enhanced structural feature representation. Subsequently, we construct a DDI network graph based on verified DDI data from authoritative databases. To alleviate the over-smoothing issue, we construct a high-order interaction topology graph via the Personalized PageRank algorithm, which enhances the aggregation of multi-hop neighborhood information and thereby improves the model's capacity to capture latent interactions among non-first-order neighbors. These three graphs are processed through shared-parameter GCNs to derive similarity and interaction embeddings. Finally, multi-head attention mechanisms uncover complementary patterns across view-specific embeddings, integrating multi-perspective features to enhance prediction performance. Rigorous evaluation across three pharmacological datasets confirms IHM-DDI's consistent superiority over contemporary methods, with a notable 3.8% F1-score gain on the ZhangDDI benchmark.