Attention Based Molecule Generation via Hierarchical Variational Autoencoder

Molecule generation is a task made very difficult by the complex ways in which we represent molecules computationally. A common technique used in molecular generative modeling is to use SMILES strings with recurrent neural networks built into variational autoencoders - but these suffer from a myriad of issues: vanishing gradients, long-range forgetting, and invalid molecules. In this work, we show that by combining recurrent neural networks with convolutional networks in a hierarchical manner, we are able to both extract autoregressive information from SMILES strings while maintaining signal and long-range dependencies. This allows for generations with very high validity rates on the order of 95% when reconstructing known molecules. We also observe an average Tanimoto similarity of .6 between test set and reconstructed molecules, which suggests our method is able to map between SMILES strings and their learned representations in a more effective way than prior works using similar methods.