Validity Improvement in MolGAN-Based Molecular Generation

Designing molecules that have desired properties is one of the challenging tasks of drug design. Among the many molecular generative models, a generative adversarial network (GAN), is able to generate molecule structures with desirable chemical properties via reinforcement learning. Generating valid molecules is the foremost task of any molecular generative model, since invalid molecules cannot be synthesized. We base our research on a molecular generative adversarial network (MolGAN) architecture to investigate how the validity score is influenced in different scenarios. First, we verify that the Vanilla GAN structure can produce valid molecules in measure, and that the reward network, along with Vanilla GAN, can further increase the validity score in a reinforcement learning manner. Then, the procedure for solely optimizing the validity score is tested, followed by an assessment of validity score maintenance while other chemical properties are being optimized. We found that multiple aspects, including loss functions, hyper parameters, and training sequences, must be carefully considered and optimized to raise the validity score of molecular generation alone or in concurrence with the optimizing of other chemical property scores.