Skip to main content

Table 6 The average test performance of the SSL-GCN models with different similarity levels of unlabeled subsets (close, normal, far) on the 12 prediction tasks in 5 repeated experiments

From: Chemical toxicity prediction based on semi-supervised learning and graph convolutional neural network

Tasks SSL-GCN (all) SSL-GCN (close) SSL-GCN (normal) SSL-GCN (far)
AUC Std AUC Std AUC Std AUC Std
NR-AR-LBD 0.7647 0.0279 0.7353 0.0353 0.7410 0.0210 0.7726 0.0242
NR-AR 0.7967 0.0251 0.7398 0.0594 0.7389 0.0401 0.7351 0.0357
NR-AhR 0.8303 0.0055 0.8261 0.0076 0.8292 0.0080 0.8278 0.0055
NR-Aromatase 0.7337 0.0057 0.7318 0.0082 0.7222 0.0131 0.7382 0.0145
NR-ER-LBD 0.6870 0.0282 0.6731 0.0261 0.6532 0.0207 0.6609 0.0253
NR-ER 0.7188 0.0196 0.7214 0.0087 0.7108 0.0133 0.7190 0.0107
NR-PPAR-gamma 0.7614 0.0212 0.7435 0.0493 0.7538 0.0164 0.7493 0.0171
SR-ARE 0.7297 0.0080 0.7308 0.0066 0.7099 0.0118 0.7172 0.0081
SR-ATAD5 0.7175 0.0143 0.6896 0.0261 0.6855 0.0295 0.7095 0.0113
SR-HSE 0.7822 0.0097 0.7833 0.0116 0.7700 0.0071 0.7745 0.0075
SR-MMP 0.8120 0.0075 0.8096 0.0097 0.8099 0.0091 0.8080 0.0091
SR-p53 0.7518 0.0198 0.7159 0.0208 0.7417 0.0129 0.7279 0.0113
  1. For comparison, the best results of the SSL-GCN models trained with the entire unlabeled dataset (all) are shown. The complete test performance can be found in the Additional file 1
  2. The bold number denotes the best result among all models (all, close, normal, far) in the corresponding task, the underlined number represents only the best result among models using different similarity levels of unlabeled subsets (close, normal, far)