Ligand Iterative Sampling for Affinity Refinement and Drug Discovery (LISARDD)

De novo drug generation is a challenging task that aims to generate novel molecules with specific properties from scratch. Deep learning can accelerate this process by efficiently exploring the drug-like chemical space. Here, we introduce LISARDD, a Reinforcement Learning framework to optimize sampling in the latent space of a pretrained target-agnostic generative model. We demonstrate that our approach can generate candidate molecules that simultaneously optimize multiple drug properties, including target-specific binding affinity, drug-likeness, and synthetic accessibility. This fully modular framework can leverage any molecular generative model, binding affinity scoring model, or optimization algorithm to identify novel drug candidates for future experimental validation.