MGRNN: Structure Generation of Molecules Based on Graph Recurrent Neural Networks
Xin Lai, Peisong Yang, Kunfeng Wang, Qingyuan Yang, Duli Yu
Abstract
Molecular structure generation is a critical problem for materials science and has attracted growing attention. The problem is challenging since it requires to generate chemically valid molecular structures. Inspired by the recent work in deep generative models, we propose a graph recurrent neural network model for drug molecular structure generation, briefly called MGRNN (Molecular Graph Recurrent Neural Networks). MGRNN combines the advantages of both iterative molecular generation algorithm and the efficiency of the training strategies. Moreover, MGRNN shows: (i) efficient computation for training; (ii) high model robustness for data; and (iii) an iterative sampling process, which allows to use chemical domain expertise for valency checking. Experimental results show that MGRNN is able to generate 69 % chemically valid molecules even without chemical knowledge and 100 % valid molecules with chemical rules.