Fig. 1

Schematic view of a typical cheminformatics workflow involving a DNN. First, a data set of compound structures and their measured activities on the desired target molecule (most often a protein) is compiled and encoded to suitable representation. Second, the encoded data is used as input of the neural network in forward mapping. A large number of architectures can be used with recurrent neural networks (RNNs) and convolutional neural networks (CNNs) as the most popular examples. Finally, the neural network is trained by backpropagating the error of a suitable loss function to adjust the activations inside the network so that the loss is minimized. Depending on the architecture, the network is trained either as a bioactivity predictor (e.g. a QSAR model) or as a molecular generator