Deep Learning-based Drug-target Interaction Prediction Study

New drug design is expensive,time-consuming,and often accompanied by safety concerns,making drug repurposing a viable option.Drug repurposing can effectively reduce the time and cost required for the development of new drugs by finding new uses for approved drugs.In order to effectively utilize existing drugs,it is significant to be aware of which proteins are targeted by which drugs,and exploring whether existing drugs can interact with new targets — Drug-Target Interaction(DTI)—has become an important topic in new Drug research and development.Identification of new DTIs is a key step in drug development,and Drug-Target Binding Affinity(DTA)reflects the reflects the intensity of drug-target Binding.Traditionally,high-throughput screening experiments for detecting biological activity between drugs and targets have been characterized by low accuracy and high cost,making this method not widely available in reality.Driven by intelligent information technology,the use of statistical methods and machine learning models to assess the interaction and binding affinity of existing drug-target protein pairs based on interaction data already measured in clinical experiments is an important alternative approach,and deep learning models are among the best-performing models for drug-target prediction problems.Therefore,in this paper,deep learning is used to predict drug-target interaction and drug-target binding affinity.The main work of this paper is as follows:(1)A new improved denoising Auto Encoder Drug-target interactions(D-DTI)algorithm is proposed to predict drug-target interactions.Noise is first added to the original DTI dataset to obtain a corrupted version of the full interaction set,and then the DAE is used to reconstruct the full input by learning potential features from the corrupted version of the dataset.However,unlike the existing DTI prediction methods using DAE,this algorithm links drugs and targets as inputs rather than separating them.To further improve the performance of the model in identifying drug-target interaction pairs,a new nonlinear computational method is added to the model to calculate the similarity information between drug-drug and target-target.By performing relevant experiments on four real datasets,namely Enzymes,Ion channels,GProtein-Coupled Receptors(GPCRs)and Nuclear receptors,the algorithm was demonstrated to outperform other baseline methods in terms of two evaluation metrics,namely AUC and AUPR.(2)A drug-target binding affinity algorithm(Graph Optimization drug-target Affinity,GODTA)based on hybrid deep neural network was proposed to capture the structural information of drugs and proteins.Different from other methods,this paper first transformed drug compounds into molecular maps,then used GCN and CNN to learn the drug features and protein sequence features represented in the maps respectively,and then used the feature optimization module to splicing the output feature vectors of each GCN layer with the final output feature vectors of CNN network.Finally,weights are allocated and added to the spliced feature vectors to obtain the final feature representation.The proposed method was validated on three benchmark drug-target binding affinity datasets,Davis,Kiba and Bingding DB,and compared with the existing state-of-the-art models in this field to demonstrate the effectiveness of the proposed algorithm in predicting affinity in terms of two evaluation metrics,CI and MSE.