A mutual-guided co-attention mechanism and heterogeneous attribute graph-based framework for drug-drug interaction event prediction

The combined use of multiple drugs helps alleviate patient resistance and enhances therapeutic efficacy. Nevertheless, this treatment strategy can also result in adverse side effects, which may compromise patient safety. Therefore, identifying potential drug-drug interactions (DDIs) and investigating their underlying mechanisms are of great significance. Existing methods predominantly predict whether drug pairs interact or whether drug-drug interaction events (DDIEs) occur, while few studies aim to reveal the specific risk levels of DDIEs, which are crucial for developing clinical medication strategies and personalized therapies. Based on this, we propose a DDIE risk level prediction method, named MCAHG-DDI, which integrates a mutual-guided co-attention mechanism with heterogeneous attribute graph learning. Specifically, we integrate the heterogeneous attribute graph with the SMILES sequences of drugs, leveraging a mutual-guided co-attention mechanism to extract the initial features of the drugs, which are subsequently input into a heterogeneous graph convolution network and a heterogeneous edge convolution network for advanced learning. Finally, we design a gated fusion mechanism to obtain the final embedding representations of the drugs. Experimental results demonstrate that MCAHG-DDI outperforms the baseline models in both binary and multi-class classification tasks. Ablation studies and case analyses further validate the superiority of the proposed model.