Study Of Drug-protein Relationships With Attention-based Deep Models

The identification of drug-target relationships(DTRs),which includes drug-target interactions(DTIs)and drug-target affinities(DTAs),is the key to modern drug discovery and drug design.It also provides direction and valuable reference for the study of drug side effects,drug repositioning,and personalized treatment.However,the number of compounds and proteins is massive,and in vitro screening experiments are labor-intensive,expensive and time-consuming with high failure rates.The computational DTR prediction methods have been paid more and more attention by researchers because of its advantages such as shortening the drug development time,reducing the blindness of new drug development,and reducing the cost of research and development.At present,there are many computational methods for predicting DTRs,mainly including similarity searching,molecular docking technology and machine learning methods.However,these methods are based on chemical and biological expertise and have their own limitations.The results are often unsatisfactory when processing complex and diverse biological information data.At the same time,with the rapid and continuous development of computer science,related databases and computer hardware conditions,deep learning has been widely used in the field of cheminformatics,providing new possibilities for identifying DTRs.The specific research contents are as follows:(1)On the prediction of drug-target interactions(DTIs),we propose an end-to-end deep learning model based on graph attention convolution neural network and cross-attention mechanism.The model takes the graph structure of the drug and the amino acid sequence of the protein as input,and uses the subgraph construction method and "word" coding to strengthen the representation of the drug and protein,respectively.The model uses graph attention convolution model to extract the drug feature matrix,which uses the attention mechanism to dynamically calculate the relationships between drug atoms and one-dimensional convolutional neural network to extract the protein feature matrix.The model obtains the attention score between the two feature matrices through the cross-attention mechanism,thereby highlighting the importance of the drug substructure and amino acid subsequence,and improving the classification ability of the model.Finally,we analyze the biological significance of the model by visualizing the attention score.(2)On the prediction of drug-target affinities(DTAs),we propose an end-to-end deep learning model ??based on convolutional neural networks and collaborativeattention mechanisms.The model takes the SMILES sequence of a drug and the amino acid sequence of a protein as inputs,uses two different one-dimensional convolutional neural network modules to extract the feature matrices of drugs and proteins,respectively,and strengthens the importance of different semantic information in the model through the collaborative attention mechanism to improve the model’s ability.Finally,we analyze the biological significance of the model by visualizing the attention score.