Attention-based Conditional VAE for Lung Cancer Drug Generation

Lung cancer attributes to one of the most prevalent types of cancer. This study proposes a conditional variational autoencoder with a self-attention mechanism to generate novel lung anticancer drug molecules. The model is built by a set of 2D convolutional layers in an encoder followed on a convolutional-based self-attention layer (SAL), while an attention Recurrent Neural Network (RNN) is used for the decoder. The experiment provided insights that the attention mechanism learns complex and long-range Simplified Molecular Input Line Entry System (SMILES) strings comparable to simple Conditional Variational Auto Encoder (CVAE). The results show that molecules generated by attention-based models were more novel and syntactically better. It was also observed that using these models for screening public databases for bio- and chemo-similar drug molecules with user-defined properties, the molecules found by the proposed model have a small mean-square error when compared to the target properties.