Classification And Activity Prediction Model Of Androgen Receptor Agonists And Antagonists Based On Graph Attention Mechanism

Androgen receptor（AR）agonists and antagonists are the main environmental endocrine disruptors.Screening and eliminating AR agonists and antagonists in the environment can effectively avoid their adverse effects on the ecological environment and organisms.At present,computer models have made great achievements in evaluating the biological activity of compounds instead of traditional experiments,but traditional machine learning cannot accurately map the complex relationship between molecular structural features and their activities.In this article,a classification model for predicting whether a compound is an AR agonist or antagonist was constructed based on the graph attention mechanism,and a regression model for predicting the activity value(p AC₅₀)of a compound as an AR agonist or antagonist was established.First,the compound molecules are encoded into molecular graphs,molecular features are extracted through the atomic and molecular embedding attention layer,task training and prediction are performed by the fully connected layer,and finally the prediction results of different tasks are output.Due to the unbalanced proportion of positive and negative samples in the classification data set,the weighted random sampling method is used to improve this problem to improve the accuracy of the classification model.The activity classification models for AR agonists and antagonists both showed excellent predictive performance on the external validation set,with AUC（area under the ROC curve）of 0.9195 and 0.9137,and Accuracy of 96.71%and 92.19%,respectively.The Pearson correlation coefficients（Rp）of the predictive regression model on the test set for the activity values of AR agonists and antagonists were 0.7852 and 0.7495,the MAE（mean absolute error）were0.3317 and 0.3850,the RMSE（root mean square error）were 0.5421 and 0.5117,The model shows better fitting and generalization.In addition,the active substructures of the compounds were characterized based on the atomic attention weight values output by the regression model.For steroidal compounds,the active structure of the model focuses on the side chains and functional groups of the steroidal nucleus;for non-steroidal compounds,the model gives high attention to structures such as phenylurea,amide,trifluoromethyl,and pyrazole value.These conclusions provide important guidance for further research on the chemical structure of endocrine disruptors.In summary,the following innovative work was carried out in this study:（1）The graph attention mechanism was used to establish an AR agonist and antagonist activity classification model with excellent predictive performance,as well as a regression model for activity value prediction.（2）The active substructures of AR agonists and antagonists are characterized based on the interpretability of the graph attention mechanism.It provides a powerful calculation method for accurately judging whether a compound has AR agonistic or antagonistic effects,and predicting its activity potential.