Fig. 1From: De novo design and bioactivity prediction of SARS-CoV-2 main protease inhibitors using recurrent neural network-based transfer learningOverview of the basic concepts showing an unfolded recurrent neural network. Molecules are represented as SMILES strings in order to train a chemical model. The SMILES string is split into individual tokens representing atoms and special environments (e.g., charged groups and stereochemistry). The tokenized molecule is then used as input to a recurrent neural network (RNN). At each time step t, the model receives as input a token and the hidden state of the previous step (ht−1). It then updates its own hidden state ht, and outputs the next token in the sequence (yi)Back to article page