Comparison of the performance of drug side effect prediction models, Random Forest, Extra Trees, Logistic regression and Neural Network when using different input drug perturbed differential gene expression profile features on two different datasets, SIDER and FAERS.

The evaluation metrics are macro-ROC-AUC and micro-ROC-AUC. In the input column, the predicted DGX means that the input feature is predicted by MultiDCP model and the experimental DGX means that the input feature is the profile collected by experiment but only the unreliable ones are used.