HierVAE++: An Update of Hierarchical Generation of Molecular Graphs Using Structural Motifs
Yadong Hu, Yue Hu, Evan Cen · 2021
Molecular generation techniques are increasingly being used for drug discovery. Previous graph generation methods utilized relatively large motifs extracted from thousands of existing molecules, limiting their effectiveness for molecular diversity. In this paper, we propose an updated version of hierarchical and graphical encoder-decoder-based method, called HierVAE++, that employs relatively smaller and more diverse graphical motif patterns as the basic blocks. The activation function is changed to a Leaky ReLU in the encode, and the hyperparameters of the neural network are adjusted accordingly. We evaluate our model in terms of the chemical properties, physical structures, the similarities of source molecules, etc. The results show that our HierVAE++ is more advantageous in generating structural diverse molecules which remain high average chemical property scores. This provides a new thinking of the drug generation industry.