Objective This study aimed to achieve molecular descriptor screening ERα regression prediction and ADMET classification prediction of antagonistic breast cancer drug candidates by using artificial intelligence algorithms to assist in drug design. Methods To address the screening problem of breast cancer drug candidates starting from the biological activity of the compounds to inhibit the target of breast cancer and their ADMET properties a two-stage screening was performed based on the obtained data of 1 974 compounds with sparse Bayesian learning and random forest algorithms respectively to obtain the top 20 molecular descriptors with the most significant effect on biological activity without strong correlation subsequently based on the screened data and its PIC50 value a QSAR model was established and the biological activity of the new compound was predicted based on the nuclear extreme learning machine algorithm optimized by the gray wolf and 11 common machine learning algorithms were compared horizontally. The ADMET classification model was constructed. Results The results show that the GWO-KELM model has higher goodness of fit and lower mean square error. The prediction accuracies of the four models were maintained above 90%. Conclusion The proposed models can effectively analyze and predict the properties of compounds which can provide a reference for the development of anti-breast cancer drug candidates.