MacGAN: A Moment-Actor-Critic Reinforcement Learning-Based Generative Adversarial Network for Molecular Generation

Deep generative models such as variational autoencoders (VAEs) and generative adversarial networks (GANs) have demonstrated significant efficacy in drug discovery. However, GANs are typically employed to process continuous data such as images and are unstable in performance for discrete molecular graphs and simplified molecular-input line-entry system (SMILES) strings. Most previous studies use reinforcement learning (RL) methods (e.g., Monte Carlo tree search) to solve the above issues. However, the generation task is time-consuming and cannot be applied to large chemical datasets due to the extensive sampling required to generate each atomic token. This study introduces a moment-actor-critic RL-based GAN (MacGAN) for the novel molecular generation with SMILES strings. MacGAN leverages the robust architecture of GAN while incorporating a simple reward mechanism, making it suitable for larger datasets compared to computationally intensive Monte Carlo-based methods. Experimental results show effectiveness of MacGAN.