Deep Learning Models for Generation of Novel Compounds

The application of deep learning models in computational chemistry and drug discovery has achieved impressive results in the past two decades. In both academia and the pharmaceutical industry, the advantages of deep neural networks have received significant attention. In this paper, we demonstrate the performance of several deep gener- ative models in drug discovery in recent years. Compared with using a model alone, a model with a hybrid structure can maximise its strengths and avoid its weaknesses. Variational auto-encoders (VAEs) are accom- panied by invalid outputs due to the posterior collapse problem. At the same time, generative adversarial networks (GANs) often struggle in a low-diversity quandary owing to GANs with continuous variables. Ad- versarially regularised graph auto-encoder (ARGA) is a graph-based pro- duction model that combines VAE and GAN. We optimised the ARGA model for molecular graph generation and found that its validity, unique- ness and novelty are superior to traditional models. We also integrated the advantages of the Junction tree autoencoder and ARGA to obtain a new model JT-ARGA. The idea of the junction tree is to improve the chemical validity of molecule generation by encoding or decoding the sub-graph of the molecule instead of creating molecules atom by atom.