Prediction of Anti-Breast Cancer Drugs Activity Based on Bayesian Optimization Random Forest

Anti-breast cancer drugs can inhibit the over-expression of estrogen receptor alpha (ERa), which is closely linked to the development of breast cancer. As such, predicting the activity of these drugs is a crucial step in anti-breast cancer drug research. To improve prediction efficiency and accuracy, this paper combines the random forest regression model with Bayesian optimization which outperforms other methods in automatic tuning of model hyperparameters to predict the activity of anti-breast cancer drugs. The preprocessing of activity and molecular descriptors data of 1974 compounds is conducted using correlation analysis and outliers elimination, and then the data are divided into training and test sets. The mean absolute error (MAE) of the model over the test sets is found to be 0.576. Additionally, the variable importance values of molecular descriptors are identified. The results of this paper show that the Bayesian optimization random forest model proposed has better prediction performance than the other three models, with mean absolute errors of 0.607, 0.605 and 0.581, respectively.