Fig. 2

The structure of NN with attention. Physicochemical and structural descriptors were calculated from the compound. Then, both features were concatenated and took into account the fully connected layer. An attention mechanism took them as the input and output an attention weight. The attention weight was element-wise multiplied with the input vector and fed into the fully connected layer. After that, the sigmoid activation function classified the DILI label. Finally, the substructures of the molecules that influenced the model prediction were identified by analyzing the attention weight