GVA: Gated Variational Autoencoder for de novo molecule generation

Molecule generation refers to the process of designing new chemicals with certain chemical properties and then optimising these properties. Following prior research, we encode chemicals as continuous vectors and decode the embedding vectors back into molecules using the variational autoencoder architecture. The encoder and decoder of the proposed variational atoencoder are based on gated recurrent unit cells. The gated recurrent unit cells limit the amount of learnable parameters in the variational autoencoder. The variational autoencoder based on gated recurrent units provides validity of 92.32 % and reconstruction accuracy of 89.63% percent, which is superior to other state-of-the-art techniques. The proposed model is effective in generating compounds with diverse properties.