DrugGen

A generative model trained on every protein-ligand pair in the public databases learns what binds, not what becomes a medicine. DrugGPT, the GPT-2 ligand generator DrugGen starts from, will emit a SMILES string for any protein sequence handed to it, but roughly one output in twenty is not a parseable molecule at all, and the valid ones carry no particular bias toward the stability, absorption, or safety characteristics that separate an interesting binder from a drug. DrugGen's premise is that approved drugs share intrinsic properties a generator can be taught to prefer.

The model is built in two stages on top of the DrugGPT checkpoint. It is first fine-tuned on a hand-curated set of drug-target pairs in which every small molecule has been approved by at least one regulatory body, then optimized with proximal policy optimization (PPO) against a reward combining a predicted protein-ligand binding affinity with a hard penalty on chemically invalid output. A separate multiplicative penalty fires when the model reproduces a molecule already in the approved set, which keeps generation from collapsing onto known drugs while still allowing them to reappear — the behavior that makes the same model usable for repositioning. DrugGen was developed at Isfahan University of Medical Sciences and published in Scientific Reports.

#Key Features

• Approved-drug conditioning: Supervised fine-tuning uses only small molecules cleared by at least one regulator, so the generator inherits the chemistry of compounds that survived development rather than of compounds that merely bind.
• Affinity-driven reinforcement learning: PPO scores each generated SMILES with a binding affinity predicted by PLAPT, a pretrained protein-ligand affinity transformer. The reward model runs during training only; nothing is optimized at inference.
• Built-in validity assessor: A custom RDKit checker flags atom-count, valence, and parsing failures and assigns them zero reward, which pushes valid structure generation to essentially every sample.
• Repurposing without collapse: A 0.7 multiplicative penalty on regenerating an approved molecule trades off fresh scaffolds against the deliberate rediscovery that drug repositioning depends on.
• Sequence-in interface: Inference accepts a raw amino acid sequence or a UniProt ID through a CLI, a Python API, or a hosted demo, all against the one published checkpoint.

#Technical Details

DrugGen is a 126.6M-parameter GPT-2 model that reads a protein sequence and a separator token and continues the string as SMILES, using DrugGPT's 53,083-token byte-pair tokenizer with inputs padded or truncated to 768 tokens. Training data comes from 1,660 DrugBank small molecules paired with 2,093 UniProt targets — 29,827 sequence-SMILES pairs — after a physician and a pharmacist reviewed 117 withdrawn drugs and excluded 50 on safety grounds. Supervised fine-tuning ran five epochs over 9,398 strings at a learning rate of 5e-4, with the third checkpoint carried forward; PPO then ran 20 epochs over the 2,053 targets short enough to fit the context window, generating 30 unique molecules per target per epoch on a single Tesla V100.

Evaluation covered eight targets: ACE and PPARG, both linked to diabetic kidney disease, plus six proteins with no approved small molecule. Against DrugGPT, validity rose from 95.45% to 99.90% and median predicted affinity from 5.81 [4.97-6.63] to 7.22 [6.30-8.07], while diversity fell from 84.54% to 60.32% and novelty from 66.84% to 41.88% — the cost of narrowing generation toward approved chemistry. Glide XP blind docking against crystal structures placed generated ligands in the reference binding site for FABP5, NAMPT, and GALM, with scores of -9.537 and -8.399 for FABP5 against palmitic acid's -6.177, and -8.381 for NAMPT against daporinad's -8.300 using a structurally distinct pharmacophore.

#Applications

The model suits medicinal chemists and computational discovery groups looking for a chemical starting point against a target that has none — six of the eight evaluation targets fell in that category. Because it returns a batch of distinct SMILES per sequence, it fits early hit generation, scaffold hopping toward new pharmacophores at an established active site, and drug repositioning, where the regeneration penalty lets approved molecules resurface for new targets.

#Impact

DrugGen is a peer-reviewed demonstration that reinforcement learning against a predicted-affinity reward measurably improves a modest GPT-2 molecular generator, and the full stack — weights, inference code, curated dataset, and demo — is public. Only the software is openly licensed, though: the weights and code are GPL-3.0, while the curated drug-target dataset is CC BY-NC 4.0 and both write-ups are non-commercial too — the Scientific Reports article CC BY-NC-ND 4.0, the arXiv preprint CC BY-NC-SA 4.0 — so commercial work can run the model but cannot retrain it on the released data. The reported affinity gains are predicted by PLAPT and corroborated by docking; no experimental validation of the generated molecules has been published. The authors note that the model cannot be directed at a specific pocket on a multi-site protein such as ACE, and that its outputs are starting scaffolds needing optimization rather than finished candidates. The same group extended the approach in DrugGen 2, which adds disease context to the target conditioning.