Reviews: Constrained Graph Variational Autoencoders for Molecule Design

Summary: This paper describes a model for generating graph-structured data, with molecule generation being the example task. This model is based around a variational autoencoder whose encoder/decoder are designed to handle graph-structured data. The decoder builds a graph sequentially by starting from an arbitrary node and sampling edges to other nodes, which are placed in a queue; upon sampling an edge to a "stopping node," the next node is taken from the queue and the process continues until there are no more nodes to expand. The distributions from which these samples are taken are a function of the graph state (notably, not the specific steps taken to arrive at the current state), where the state vectors are encoded using a gated graph neural network (GGNN). Additionally, masking functions can be specified that serve as hard constraints on the sorts of edges that may be sampled (in case these would lead to graphs that are disallowed, e.g. that would lead to impossible molecules).