arxiv

DrugGen 2: A disease-aware language model for enhancing drug discovery

Ali Motahharynia, Mohammadreza Ghaffarzadeh-Esfahani, Mahsa Sheikholeslami, Navid Mazrouei et al.Jul 9, 2026Original

ResearchPod Summary

Research Question and Approach

Drug discovery is often limited by a focus on target-based design that ignores the broader disease context. Because the same protein target can have different roles depending on the disease state, the authors developed DrugGen-2, a generative model that incorporates both target protein sequences and disease ontology (using MeSH hierarchies) to design small molecules. The model was built by fine-tuning a pre-trained GPT-2 architecture using a two-step process: supervised fine-tuning on a curated dataset of approved drugs, followed by reinforcement learning via Group Relative Policy Optimization (GRPO). The training was guided by reward functions that prioritized chemical validity, structural novelty, diversity, and high predicted binding affinity.

Main Findings

DrugGen-2 was evaluated against baseline models (DrugGPT and DrugGen) across five protein targets relevant to diabetic nephropathy. The model demonstrated superior performance in several key metrics:

Generation Capacity: It consistently generated a higher number of unique molecules compared to the baselines.
Chemical Quality: The model achieved near-perfect structural validity (99–100%) and produced molecules with higher structural similarity to known approved drugs.
Binding Affinity: Predicted binding affinities were significantly higher across all tested targets. Molecular docking simulations confirmed that several generated ligands achieved binding scores superior to reference drugs like enalapril.

Why It Matters

By integrating disease-specific context into the generative process, DrugGen-2 addresses a critical gap in AI-assisted drug discovery. This approach allows for the design of molecules that are not only optimized for a specific protein target but are also tailored to the biological requirements of a particular disease. This capability is particularly valuable for drug repurposing and de novo design, potentially accelerating the transition from computational hit identification to clinical development by providing more contextually relevant leads.

Paper Details

arXiv: 2607.08404

Topics

DrugGen-2Disease-aware drug discoveryReinforcement Learning (GRPO)Molecular DockingDrug RepurposingDisease-conditioned drug designLarge Language Models (LLMs)Group Relative Policy Optimization (GRPO)Supervised Fine-Tuning (SFT)MeSH Directed Acyclic Graph (DAG)Binding AffinitySMILES

Podcast transcript

Alex: Welcome to another episode of ResearchPod. Today we're looking at a system called DrugGen-2—an AI designed to help create medicines by considering not just the target protein, but the broader disease environment surrounding it.

Sam: So the paper's argument is that we've been designing drugs in a kind of vacuum—focusing only on the protein while ignoring everything around it?

Alex: Exactly. Think about it this way: a protein in a healthy cell and the same protein in a diseased cell don't behave identically. The disease changes the environment, which changes how the protein acts. So a drug designed only around the protein's general shape might not work as well as one tailored to that specific disease setting.

Sam: It's like fixing a traffic jam—you need to know whether it's caused by an accident or a parade before you can pick the right detour. How does DrugGen-2 actually bring that disease information in?

Alex: They use something called MeSH—Medical Subject Headings. Think of it as a structured library catalog for all of medicine. It organizes diseases, symptoms, and treatments into a hierarchy, showing how concepts relate to one another. So the system doesn't just know "this is a kidney disease." It understands where that disease sits in the broader map of medicine.

Sam: And once it has that map, how does it actually generate a drug molecule?

Alex: It uses a Large Language Model—the same kind of technology that predicts the next word in a text message. But instead of predicting words, it predicts the next piece of a chemical structure, one building block at a time. By feeding it both the disease map and the protein's sequence, the model learns to associate specific chemical patterns with the needs of that exact disease environment.

Sam: That's a useful way to think about it. But how do they stop it from generating chemical nonsense—structures that look plausible on paper but would never work in a real body?

Alex: That's where the training process becomes important. They use a technique called Group Relative Policy Optimization—GRPO for short. Imagine a teacher who gives a class of students the same problem, then grades all their answers together. The students who follow the rules—chemical stability, correct structure, strong binding to the target—get rewarded. The ones who don't, get penalized. Over many rounds of this, the model learns to consistently produce molecules that are both chemically valid and likely to attach firmly to the target protein.