Deep Learning Model-Based Classification And Prediction On Compound-Protein Interactions

Identification of drug targets plays an important role in the fields of drug discovery and biopharmaceuticals.The core of this is the identification of compound-protein interactions.Traditional experimentally-based drug development models have accumulated a large amount of basic data for compound-protein interactions.However,this model has the characteristics of complex processes,incomplete identification,long cycles,high costs,and low success rates.In recent years,deep learning models trained based on a large amount of basic data have achieved amazing results in the classification and prediction of different fields(image recognition and speech processing,etc.).The massive accumulation of basic data,the rapid development of CPU + GPU,and the rapid evolution of deep learning models have provided the possibility to classify and predict compound-protein interaction relationships based on this model.Therefore,this study uses a deep learning model to classify and predict the compound-protein interaction relationship,and can comprehensively learn the characteristics of the accumulated basic data of the compound-protein interaction relationship in a short time,and predict based on this characteristic new compound-protein interactions to provide low-cost clues for drug target identification.The experimental data used in this study is mainly from the BindingDB database.After processing the original data,we get 1224408 binding positive sample data that compounds can interact with the proteins.The classification label is set to 1.The random generation algorithm is used to generate negative sample data according to positive and negative samples 1: 2 with its label set to 0.Then,the paper uses the TensorFlow framework to build a powerful deep learning model for classifying and predicting compound-protein interactions.Based on the above experimental data,we carry out model training,parameter adjustment and determination of the winning model.The entire research process includes data collection and processing,model construction and implementation,and determination.During this period,a lot of neural network experimental solutions were tried.After comparison,the final model uses a deep neural network to extract and splice the compound and protein data separately,and it uses a recurrent neural network to classify and predict the interactions.The number of model parameters reached 39.27 million,the accuracy on the test set reached 95.82%,the F1-score value was 94.05%,and the AUC value was 98.71%.Target recognition for a certain compound has also obtained better prediction results.The experimental results of this study suggest that the method based on deep learning has certain theoretical and practical significance for the classification and prediction of compound-protein interactions.