Skip to main content

Table 6 Comparison of our CNN and DeepTox (the winning model of the TOX 21 Challenge 2014)

From: Convolutional neural network based on SMILES representation of compounds for detecting chemical motif

Input Model Ave. AR AR-LBD ER ER-LBD AhR Aromatase PPAR- γ ARE ATAD5 HSE MMP p53
SMILES Matrix CNN 0.813 0.789 0.793 0.776 0.765 0.905 0.786 0.791 0.754 0.803 0.835 0.928 0.832
ECFP DNN 0.768 0.850 0.690 0.840 0.760 0.660 0.720 0.700 0.730 0.860 0.810 0.820 0.780
ECFP+DeepTox DNN 0.837 0.778 0.825 0.791 0.811 0.923 0.804 0.856 0.829 0.775 0.863 0.930 0.860
ECFP+DeepTox SVM 0.832 0.882 0.748 0.799 0.798 0.919 0.819 0.856 0.818 0.781 0.848 0.946 0.854
ECFP+DeepTox RF 0.820 0.776 0.812 0.770 0.746 0.917 0.806 0.827 0.810 0.786 0.826 0.945 0.835
ECFP+DeepTox ElNet 0.803 0.788 0.692 0.765 0.805 0.897 0.763 0.805 0.778 0.768 0.844 0.924 0.818
  1. Our CNN takes SMILES feature matrices as input, while DeepTox uses ECFP and its original features