Drug target binding affinity prediction considering the interaction of molecule atom and protein subsequences

Abstract

The identification of drug target interactions is an important task in drug discovery. To model the interaction of drug and protein, most of the computational methods of drug target affinity prediction have been proposed the global interaction between compound and protein. We propose a new model called GraphATTDTA that could model the important relationships between molecule substructure and protein subsequences using attention mechanism. The atom of molecule is encoded with graph neural networks and the subsequence of protein with 1D convolutional neural networks. The attention mechanism would help to drug feature and protein feature reflect the mutual relationship. We evaluate out model on two drug target affinity benchmark datasets, Davis and KIBA. We compared our model with DeepDTA and GraphDTA. The results show that our model achieves improvement of prediction performance on Davis Dataset and slight lower on KIBA dataset. In case study, we illustrated the biological insights to understand the predicted binding affinity.