DrugGen 2: A Disease-Aware Language Model for Enhancing Drug Discovery

DrugGen-2 generates small molecules conditioned on both disease ontology and target-protein sequences. Built by fine-tuning GPT-2 with supervised learning followed by group relative policy optimization (GRPO), it rewards chemical validity, novelty, diversity, and predicted binding affinity. On five targets associated with diabetic nephropathy, the authors report improvements over DrugGPT and DrugGen in uniqueness, similarity to approved drugs, and predicted affinity. Docking identified candidates with reported affinities as strong as -9.917, compared with -8.283 for enalapril against ACE. These results are computational and do not establish biochemical activity, selectivity, safety, or clinical benefit.