Abstract
For the purpose of developing new drugs and repositioning existing ones, accurate drug-target affinity (DTA) prediction is essential. While graph neural networks are frequently utilized for DTA prediction, it is difficult for existing single-scale graph neural networks to access the global structure of compounds. We propose a novel DTA prediction model in this study, MAPGraphDTA, which uses an approach based on a multi-head linear attention mechanism that aggregates global features based on the attention weights and a multi-scale gated power graph that captures multi-hop connectivity relationships of graph nodes. In order to accurately extract drug target features, we provide a gated skip-connection approach in multiscale graph neural networks, which is used to fuse multiscale features to produce a rich representation of feature information. We experimented on the Davis, Kiba, Metz, and DTC datasets, and we evaluated the proposed method against other relevant models. Based on all evaluation metrics, MAPGraphDTA outperforms the other models, according to the results of the experiment. We also performed cold-start experiments on the Davis dataset, which showed that our model has good prediction ability for unseen drugs, unseen proteins, and cases where neither drugs nor proteins has been seen.