ExSelfRL

Efficiently searching for novel molecules with specific properties is critical to molecular generation. Some existing works focus on combining deep generative models and reinforcement learning to generate molecules with targeted properties, but there is still the problem of reduced model effectiveness due to sparse rewards. To address the problem, an exploration-inspired self-supervised reinforcement learning (ExSelfRL) method for molecular generation is proposed. By constructing an exploration-inspired reward-shaping framework, ExSelfRL can effectively mine the intrinsic rewards in the molecule generation process based on novelty. Then, driven by intrinsic and primitive sparse rewards, ExSelfRL establishes a self-supervised reinforcement learning agent capable of exploring a broader chemical space to find molecules with better properties. In addition, a dominant set of molecules is defined from the sampled molecules that can further improve their property scores. The experimental results illustrate that ExSelfRL can generate molecules with higher property scores than existing methods.