Skip to main content


Fig. 2 | Journal of Cheminformatics

Fig. 2

From: QBMG: quasi-biogenic molecule generator with deep recurrent neural network

Fig. 2

Network architecture and training procedure. a Unfolded representation of the training model, which contains embedding layer, GRU structure, fully-connected linear layer and output layer. The structure of GRU cell is detailed on the right. b Flow-chart for the training procedure with a molecule. A vectorized token of the molecule is input as \(x_{t}\) in a time step, and the probability of the output to \(x_{t + 1}\) as the next token is maximized. c The new molecular structure is composed by sequentially cascading the SMILES sub-strings replied by the RNN network

Back to article page