MoFlowGAN: Combining adversarial and likelihood learning for targeted molecular generation

Deep generative models for molecular graphs offer a new avenue for property optimization in drug discovery. Optimizing differentiable models that generate molecular graphs is certainly faster, cheaper, and much more accessible than traditional methods of chemical synthesis. Recent advances in generative modeling have managed to address many of the challenges surrounding generation of chemically-valid molecular graphs from latent representations, however the question of generating high-quality molecules remains. Herein we introduce MoFlowGAN a tandem normalizing flow model that can also be trained on both adversarial and reward objectives. We train our model on QM9 to generate high-quality and drug-like compounds. Our experiments show that MoFlowGAN is competitive with current state-of-the-art generative models while requiring far fewer training resources.