Drug-Target Affinity Prediction Based on Dynamic Graph Isomorphism Network and Multi-Scale Features
Huaizhou Yang, Xiaohan Tong, Nan Ma, Xin W. Chen · 2024
The prediction of drug target affinity plays a crucial role in the field of drug development. To address the limitations of existing prediction methods, which often overlook the two-dimensional structural information of drug molecules and lack deep feature learning, we propose a novel model for drug target affinity prediction called GDB_DTA. This model incorporates a dynamic graph isomorphism network and multi-scale features. Firstly, it encodes the graph representation of the drug molecule using a multi-layer graph isomorphism neural network combined with a dynamic graph attention network. Secondly, it encodes the target sequence using both BiLSTM and 1D-CNN models. Subsequently, by concatenating these encoded features, fusion features are obtained and fed into a fully connected layer to predict the affinity score accurately. Experimental results demonstrate that our proposed model outperforms six mainstream methods on both DAVIS and KIBA datasets, significantly enhancing prediction accuracy.