Classification of thyroid hormone receptor agonists and antagonists using statistical learning approaches

In silico models are presented for modeling and predicting thyroid hormone receptor (TR) agonists and antagonists. A data set consisting of 258 compounds is used in the present work. The C4.5, random forest (RF) and support vector machine (SVM) statistical methods were used for evaluation. The performance of the quantitative structure-activity relationships was further validated with fivefold cross-validation and an independent external test set. The C4.5 model is slightly weak, and the prediction accuracies of the agonists and antagonists are 93.2 and 57.8% for cross-validation, respectively, averaging 83.1% of correctly classified compounds in the test set. The RF model possesses an average prediction accuracy of 84.0 and 87.1% for the cross-validation and external validation, respectively. Furthermore, the overall prediction accuracy and the external prediction accuracy are 96.6 and 97.2%, respectively, for the SVM model. The results would validate the reliability of the derived models, further demonstrating that RF and SVM models are useful tools capable of classifying TR-binding ligands as agonists or antagonists.