Prediction Study Of Protein-drug/Protein Interactions Based On Graph Neural Networks

Predicting the interaction between drugs and proteins is still an important research content in the design process of new drugs,but it takes a lot of time and material cost to compare with traditional biological experiments.In recent years,with the continuous development of artificial intelligence and big data,algorithms to predict the interaction between drugs and proteins based on machine learning have gradually emerged.However,these algorithms have problems such as failure of large-scale prediction,low prediction accuracy and difficulty to obtain the required prior information.To this end,this thesis proposes the prediction algorithm of protein and drug / protein interaction based on graph neural network(GNN),with the following three contents:(1)To effectively extract the features of drugs and proteins,the Bert model and CT method were used to extract the features of proteins,and the SMILES2 Vec method to extract drugs.(2)To accurately predict drug-protein interactions,a drug-protein interaction prediction algorithm is proposed.Current relevant research methods only use drug similarity or target-protein similarity information to describe drug molecules and protein sequences,and do not fully contain important information about drug-protein interactions.To this end,the algorithm uses the above research method to first extract the features of proteins and drugs,and then use the k-mer features of proteins and the FP features of drug molecules.To verify the effectiveness of the algorithm,the relevant experimental comparison was performed.Experimental results show that the proposed algorithm is higher than the existing algorithms in the prediction accuracy of drug-protein interactions.(3)To predict protein – protein interactions,a prediction algorithm for protein – protein interaction based on graph neural network.This algorithm improves the algorithm model described above and adds an MLP layer to solve the dimensional mismatch problem.In addition,the interaction between proteins and proteins is also scored and evaluated in the algorithm.Experimental results show that the proposed algorithm is better than the existing prediction accuracy of protein-protein interaction,and the operation speed is greatly improved.