Predictability Of Interpretable Drug Target Affinity Based On Attentive Pooling Networks

In the topic of drug discovery,we can achieve the purpose of drug discovery by studying the interaction between drugs and proteins.Drug protein affinity refers to the strength of drug and protein binding after interaction.Therefore,it is very important to find out the affinity of protein and ligand drugs.In the earliest days,drug discovery could only rely on people discovering by accident.Until biology,chemistry and other disciplines gradually developed,people discovered drugs in experiments.And through the deepening of biochemical research,people can begin to design drugs with purpose.But until this period,all drug discovery abstracts could only be carried out in a way that reached clinical trials.However,the way of clinical trials was timeconsuming and expensive,and the success rate could not be guaranteed.With the development of computer science,computing methods have been introduced into the subject of drug discovery.In the prediction of drug-protein interaction affinity,the calculation method has shown excellent advantages.Not only the success rate is higher and the speed is faster,but also the economic cost is greatly reduced.The current mainstream computing methods are mainly the following: molecular-based docking method,similarity-based method and machine learningbased method.The method based on molecular docking needs to understand the threedimensional structure of small molecules of protein and drug,find the best binding site of protein and drug in three-dimensional space,and predict the binding affinity.The method based on similarity mainly relies on a hypothesis: drugs and proteins with similar sequences and structures should also have similarities in the performance of their biological functions.Machine learning methods mainly rely on computers to learn some potential features in proteins and drugs,so as to achieve the purpose of predicting the affinity of unknown proteins and drugs.However,many current machine learning methods applied to the prediction of drug protein affinity are not interpretable.Therefore,a bioexplainable model is proposed in this paper,and it proves that it can effectively predict the affinity of drug proteins to a certain extent.This paper proposes a prediction model for the affinity of drug proteins based on Attentive Pooling Networks,and the model has biological interpretability.The main feature of the model in the data processing part is the use of the "word segmentation" method,while the main feature of the network part is the use of a two-way attention mechanism.In the experimental part,two sets of experiments were set up based on two different ways of expressing proteins(amino acid sequence,domain and motif).The final experimental results showed that using domains and motifs as protein expressions can achieve better results.In addition,three sets of comparative experiments are also set up in this article.The first two groups respectively remove the "word segmentation" thoughts of the data processing part and the two-way attention mechanism thoughts of the network construction part,and compare with the results of the original experimental model.The final experimental results show that The lack of any part will make the experiment result worse.The third group compares the differences in the results of using LSTM and GRU to treat proteins and drugs in the experimental model.The results prove that the effect of using GRU is slightly better than that of using LSTM.