Table 2.

Performance comparisons for different architectures under stratified 5-fold cross-validation on CRISPOR dataset

Model	Min_AUC	Max_AUC	Mean_AUC	Var_AUC
FNN_2layer	0.852	0.891	0.842	0.010
FNN_3layer	0.963	0.977	0.970	0.005
FNN_4layer	0.951	0.960	0.954	0.009
CNN_std	0.954	0.983	0.972	0.010
CNN_nbn	0.929	0.973	0.954	0.022
CNN_nd	0.953	0.974	0.969	0.013
CNN_np	0.720	0.981	0.899	0.093
CNN_pool_win3	0.632	0.979	0.903	0.137
CNN_pool_win7	0.943	0.983	0.967	0.015

Bold values signifies: CNN_std achieved the highest Mean_AUC (0.972) and highest Max_AUC (0.983) under stratified 5-fold cross-validation in predicting off-targets among all neural network based models, FNN_3layer also accomplished the competitive performance (Mean_AUC = 0.970) with the highest Min_AUC (0.963) and the lowest AUC variance (0.005).