Graph Unseen Node Embedding Learning Network for Drug Target Affinity Prediction

The procedure of drug development is both financially burdensome and time-demanding. In order to accelerate the pace of drug discovery and development, it is worthwhile to understand how closely a drug binds to a specific target, known as drug-target binding affinity (DTA). Traditional experimental approaches to determine drug-target binding affinity usually require a considerable amount of time, so treating the task of combining target affinity with drug confirmation as a regression task captivates people’s interest. As machine learning progresses, the utilization of deep models for the prediction of drug-target binding affinity is becoming more prevalentand, people take advantage of deep learning frameworks to learn representation of ligand chemical compounds and protein sequences. Recent methods use graph structures to represent drug molecules and then use graph convolutional networks for characteristic encoding. Traditional graph convolutional networks require global structural information when training embeddings and cannot naturally generalize to unseen nodes. In an effort to find solutions to the aforementioned issues, we proposed the graph unseen node embedding learning network for drug target affinity prediction (GUNE-DTA), using SAGEConv, an improved network of graph convolution network, to perform representation learning on drugs. SAGEConv introduces a framework GraphSAGE, which proposes an unsupervised node representation learning method that can use node attribute data to effectively produce node embeddings for previously unknown data, and obtain node embeddings through neighbor sampling and aggregation. It has good scalability and cross-network generalization capabilities. We study the model’s performance on two benchmark datasets. The findings suggest that GUNE-DTA surpasses established test domain models in terms of all performance metrics.