Automatic Molecule Generation Using Manifold Guidance GAN and Genetic Algorithm

Abstract

Deep generative models used to generate molecules have demonstrated their superior performance in the creation of novel structures. However, mode collapse is a severe and frequent issue in the GAN-based molecule generation models, in which generator learns a few modes of the data distribution while ignoring others. In this work, we introduce Mol Manifold Guidance Generative Adversarial Network (Mol-MGGAN) to solve this problem. Mol-MGGAN extends generative adversarial networks by introducing a manifold guidance network, which contains a graph encoder that maps molecules into a latent manifold space that covers overall modes of the data distribution, and a discriminator that distinguishes molecules in the manifold space. The guidance network can explicitly prevent the generator from mode collapse through forcing the generator to learn the overall modes of the data. We use the genetic algorithm to further enhance the generator’s ability to produce novel and unique molecules. In the experiments on the QM9 chemical database, we demonstrate that Mol-MGGAN generates nearly 100% valid molecules. Most importantly, we generate more unique and novel molecules compared to the previous GAN-based molecule generation model. The result of the experiments shows that Mol-MGGAN reduces mode collapse during the molecular generation.