Isfahan team's DrugGen-2 ties drug design to disease context

Most generative chemistry models still condition on a protein target and hope the disease sorts itself out, which is a strange assumption given how differently the same target behaves in different diseases. A team out of Isfahan University of Medical Sciences, led by Mohammadreza Ghaffarzadeh-Esfahani and Mahsa Sheikholeslami with Ali Motahharynia as corresponding author, put out a paper on Hugging Face's daily papers listing that tries to close that gap. Their model, DrugGen-2, takes both a disease MeSH directed acyclic graph and a target amino acid sequence as input, and outputs SMILES for that specific disease-target pair.

The recipe is deliberately unflashy. They start from DrugGPT, a pre-trained GPT-2 model already specialized for ligand generation, then run six epochs of supervised fine-tuning on a curated dataset of approved drugs linked to their diseases and targets, then ten epochs of group relative policy optimization with three reward functions: predicted binding affinity from a transformer called PLAPT, batch-level molecular diversity, and novelty against the approved-drug set. GRPO is the same class of reinforcement-learning objective that has been showing up in open-model reasoning work; here it is doing the shaping instead of a bespoke RL loop.

On the benchmark side the authors picked five targets tied to diabetic nephropathy — ACE, PPAR-gamma, NOS3, PAI-1 and TGF-beta-1 — and asked each model to generate 500 unique candidates. DrugGen-2 delivered 409 to 444 unique molecules per run, versus 219 for DrugGPT and 50 for DrugGen, with validity at 99 to 100 out of 100 sampled. Similarity to approved drugs came in at 0.70 for DrugGen-2, 0.64 for DrugGen and 0.30 for DrugGPT. Median predicted binding affinities landed between 9.26 and 9.97, well above DrugGPT's 5.86 to 6.22 range and DrugGen's 7.15 to 8.49. Docking against ACE surfaced candidate ligands with predicted affinities of -9.917, -9.485 and -9.367, which the paper notes are stronger than reference drug enalapril's -8.283.

The honest caveat is that every headline number here lives inside a computer. PLAPT is a predicted score and the docking work is a simulation, so the model may partly be learning to please its own reward stack. The comparative work also runs entirely on one disease and against two prior open baselines from the same lineage; there is no head-to-head against newer target-aware generators like Mol-LLM or TamGen, and the reporting does not tell you whether the weights or the curated disease-target-drug dataset will be released, or whether any of the top candidates have been touched by a wet lab.

What is interesting even with those caveats is the direction. If a modest academic team can get disease-conditioned generation working on a GPT-2-scale base with GRPO rather than a frontier model, the compute bar for building disease-specific drug pipelines drops meaningfully, and small biotech and academic groups become plausible participants in the design step rather than just consumers of frontier-lab tools.