Multimodal Feature Fusion-Based Node Representation Learning for Drug-Drug Interaction Prediction

Predicting Drug-drug interaction (DDI) is essential for drug combination therapy to reduce unexpected adverse drug reactions. However, existing methods cannot adequately consider multimodal features or molecular graph structural information, and it is a challenge to extract and fuse multimodal features effectively. In this study, we propose a multimodal feature fusion-based node representation learning for Drug-Drug interaction prediction, termed MFF-NRL. We used a graph convolutional network to extract the features of molecular graphs in terms of a node feature update strategy. Then, we take the molecular graph embeddings as a feature matrix, construct an adjacency matrix based on the DDI network, and propose a representation learning method based on a multi-head self-attention mechanism. Finally, a multimodal feature fusion module is presented to effectively integrate the molecular graph and DDI network embedding information, and the multi-layer perceptron (MLP) layer is further applied for DDI prediction. The MFF-NRL algorithm's effectiveness has been verified by four representative evaluation metrics on the ZhangDDI dataset. The experimental results demonstrate that MFF-NRL performs better than state-of-the-art approaches.