Novel Molecule design with POWGAN, a Policy-Optimized Wasserstein Generative Adversarial Networks

This study introduces Policy Optimized Wasserstein GAN (POWGAN), a novel generative model that integrates reinforcement learning policy-driven optimization. POWGAN employs a dynamically scaled reward function that adaptively adjusts the training focus, promoting the generation of novel molecules with targeted properties, such as graph connectivity to deliver non-fragmented molecules. The results demonstrated substantial improvements over previous approaches, highlighting the model's ability to achieve nearly 100% connectivity and significantly enhance generative capacity by up to eight-fold, producing more than 10,000 novel molecules. R-MedGAN, utilizing POWGAN's capability to produce structurally diverse molecules, facilitated the exploration of novel chemical regions and substantially expanded the accessible chemical space. These findings underscore the effectiveness of adaptive reinforcement-driven strategies in generative adversarial networks oriented by rewards for molecular discovery.