DeepNC: a framework for drug-target interaction prediction with graph neural networks

PeerJ

article

Huu Ngoc Tran Tran¹  J. Joshua Thomas¹  Nurul Hashimah Ahamed Hassain Malim² 

[1] Department of Computing, UOW Malaysia, KDU Penang University College;School of Computer Sciences, Universiti Sains Malaysia

关键词: Drug-target interaction;    Binding affinity;    Drug discovery;    Deep learning;    Graph neural networks;    Cheminformatics;   

DOI  :  10.7717/peerj.13163

学科分类：社会科学、人文和艺术（综合）

来源: Inra

【 摘 要 】

The exploration of drug-target interactions (DTI) is an essential stage in the drug development pipeline. Thanks to the assistance of computational models, notably in the deep learning approach, scientists have been able to shorten the time spent on this stage. Widely practiced deep learning algorithms such as convolutional neural networks and recurrent neural networks are commonly employed in DTI prediction projects. However, they can hardly utilize the natural graph structure of molecular inputs. For that reason, a graph neural network (GNN) is an applicable choice for learning the chemical and structural characteristics of molecules when it represents molecular compounds as graphs and learns the compound features from those graphs. In an effort to construct an advanced deep learning-based model for DTI prediction, we propose Deep Neural Computation (DeepNC), which is a framework utilizing three GNN algorithms: Generalized Aggregation Networks (GENConv), Graph Convolutional Networks (GCNConv), and Hypergraph Convolution-Hypergraph Attention (HypergraphConv). In short, our framework learns the features of drugs and targets by the layers of GNN and 1-D convolution network, respectively. Then, representations of the drugs and targets are fed into fully-connected layers to predict the binding affinity values. The models of DeepNC were evaluated on two benchmarked datasets (Davis, Kiba) and one independently proposed dataset (Allergy) to confirm that they are suitable for predicting the binding affinity of drugs and targets. Moreover, compared to the results of baseline methods that worked on the same problem, DeepNC proves to improve the performance in terms of mean square error and concordance index.

【 授权许可】

CC BY