Graph Convolutional Policy Network for Goal-Directed Molecular Graph\n Generation

Generating novel graph structures that optimize given objectives while\nobeying some given underlying rules is fundamental for chemistry, biology and\nsocial science research. This is especially important in the task of molecular\ngraph generation, whose goal is to discover novel molecules with desired\nproperties such as drug-likeness and synthetic accessibility, while obeying\nphysical laws such as chemical valency. However, designing models to find\nmolecules that optimize desired properties while incorporating highly complex\nand non-differentiable rules remains to be a challenging task. Here we propose\nGraph Convolutional Policy Network (GCPN), a general graph convolutional\nnetwork based model for goal-directed graph generation through reinforcement\nlearning. The model is trained to optimize domain-specific rewards and\nadversarial loss through policy gradient, and acts in an environment that\nincorporates domain-specific rules. Experimental results show that GCPN can\nachieve 61% improvement on chemical property optimization over state-of-the-art\nbaselines while resembling known molecules, and achieve 184% improvement on the\nconstrained property optimization task.\n