We introduce a generative drug-design framework that combines large chemical language models (CLMs) pretraining, target specific masked-language fine-tuning, and reinforcement learning (RL) to create novel small molecule inhibitors of EGFR. Using a multi-objective reward that balances predicted potency, drug-likeness, synthetic accessibility, and structural novelty, the model learns to explore chemically valid and diverse regions of EGFR-relevant chemical space beyond known inhibitors. The resulting compounds exhibit improved computational binding trends relative to reference EGFR inhibitors and include highly novel chemotypes with no close analogs in the training set. This study demonstrates how integrating pretrained chemical language models with reinforcement learning can accelerate target focused de novo molecular design and provides a generalizable framework for future applications in kinase inhibitor discovery.
Chai et al. (Sat,) studied this question.