Prediction Methods Of Drug-target Interactions Based On Graph Neural Network

Drugs combine with targets to influence the physiological functions and metabolic effects of the body,and the accurate identification of such drug-target interactions(DTIs)is a fundamental and breakthrough point in the field of drug development.Computer-aided drug discovery can effectively predict interactions between drugs and targets,and can help biological researchers narrow their search space and reduce experimental exertion.Although researchers have obtained some results in this area,there are still many elements and questions that can be improved and explored.This thesis regards the task of predicting drug-target interactions as a binary classification problem,and proposes two drug-target interactions prediction models.Both of them achieve the prediction of interaction pairs through the SMILES of small molecule chemosynthetic drugs and the amino acid sequences of targets.Firstly,a word vector-based drug-target interactions prediction model in terms of the one-dimensional form of the features used is proposed in this thesis.This method treats SMILES and amino acid sequences as sentences and encodes them as word vector features.The interaction features are extracted from drug and target word vector features by a three-layer one-dimensional convolutional structure.Then the results are obtained by a feature fusion layer and a three-layer fully-connected structure.This classification prediction model for drug-target interaction pairs is implemented based on Text CNN using transfer learning idea.The experimental results show that the word vector-based prediction model achieves the best results on two benchmark datasets,demonstrating that this method has a good prediction effect on drug-target interaction pairs.Secondly,in order to better extract the features of drug-target pairs and further improve the prediction accuracy,a drug-target interaction prediction model based on graph neural network from the two-dimensional molecular graph form of drugs is proposed in this thesis.In this method,the compound is represented as a two-dimensional molecular graph,that is,SMILES is encoded as a graph structure through an algorithm.The graph representation of SMILES and the word vector of amino acid sequences are used as model inputs respectively,and the feature extraction is done through a five-layer graph isomorphic network structure and a four-layer one-dimensional convolutional structure respectively,followed by a feature fusion layer and a three-layer fully-connected structure to process the obtained interaction pairs features.Finally the one-dimensional classification prediction results of drug-target interaction pairs are obtained.The experimental results show that the graph neural network-based prediction model achieves the best results on the three benchmark datasets and outperforms the word vector-based model on all four benchmark datasets.Finally,a case study of Corona Virus Disease 2019(COVID-19)related drugs and their corresponding interacting targets was conducted in order to demonstrate the effectiveness of the models proposed in this thesis in practical applications.In case of cold start,where drugs and some of targets to be predicted were not included in the pre-trained datasets,the pre-trained graph neural network models obtained based on four benchmark datasets successfully predicted five drug-target interaction pairs of COVID-19.