Designing ligands that selectively bind to protein targets is a cornerstone of drug discovery. However, identifying suitable ligands that fit the binding site and preserve protein function remains experimentally time consuming and expensive. To accelerate de novo drug discovery, traditional inverse design methods have opened many promising avenues.1 Large language models (LLMs) have further transformed machine learning (ML) applications, shifting from purely predictive tasks to leveraging vast corpora for chemistry. Their generative capabilities have led to the development of text-to-molecule models, offering new approaches for molecular design and drug discovery. Yet, existing LLMs often struggle to generate chemically valid molecules and typically lack mechanisms for reasoning-based validation of their outputs. Much less work has focused on creating reasoning LLMs for scientific tasks,2 and relatively little has been done specifically for protein-binding tasks.
In this work we propose a reasoning LLM which is capable of designing protein-binding ligands. Through group relative policy optimization (GRPO)3 method with a verifiable docking score reward, we demonstrate how reasoning models can enable in sampling valid ligand candidates and justify their predicted binding capabilites. This reasoning driven approach helps towards an efficient and accelerated path towards drug discovery.
