A multi-scale feature fusion model based on biological knowledge graph and transformer-encoder for drug-drug interaction prediction

Drug-Drug Interaction (DDI) refers to the combined effects that occur when a patient takes multiple medications simultaneously or within the same period. This interaction can either enhance the therapeutic effects of the drugs or inhibit their efficacy, and in severe cases, it can even lead to adverse drug reactions (ADRs). Thus, it is crucial to identify potential DDIs, as this information is significant for both biological research and clinical medicine. However, most existing works only consider the information of individual drugs or focus on the local correlation between a few medical entities, thus overlooking the global performance of the entire human medical system and the potential synergistic effects of multi-scale information. Consequently, these limitations hinder the predictive ability of models. In this paper, we propose an innovative multi-scale feature fusion model called ALG-DDI, which can comprehensively incorporate attribute information, local biological information, and global semantic information. To achieve this, we first employ the Attribute Masking method to obtain the embedding vector of the molecular graph. Next, ALG-DDI leverages heterogeneous graphs to capture the local biological information between drugs and several highly related biological entities. The global semantic information is also learned from the medicine-oriented large knowledge graphs. Finally, we employ a transformer encoder to fuse the multi-scale drug representations and feed the resulting drug pair vector into a fully connected neural network for prediction. Experimental evaluations on datasets of varying sizes and different classification tasks demonstrate that ALG-DDI outperforms other state-of-the-art models.